Adjunct Proceedings

of the
18th International Conference on
User Modeling, Adaptation, and
Personalization
Posters and Demonstrations


Big Island of Hawaii, HI, USA
June 20-24, 2010
http://www.hawaii.edu/UMAP2010/



Fabian Bohnert and Luz M. Quiroga
(Poster and Demo Co-Chairs)




Additional Reviewers
Pei-Chia Chang
Alex Chen
Caterina Desiato
Curtis Ikehara
Viil Lid
Lipyeow Lim
Victor Magkikh
Michael-Brian Ogawa
Scott Robertson
Yanir Seroussi
Dennis Streveler
William Yip


Table of Contents
Posters
A Bayesian Hierarchical Preference Model for Context-Aware Recommendations ...................... 1
Hideki Asoh, Yoichi Motomura, and Chihiro Ono
WAI-WEB: Traveler assistance adapting user interface and routes to people with cognitive
disabilities ....................................................................................................................................... 4
Linda Barros, Luis Rojas, Pablo A. Haya, Rosa M. Carro, and Pilar Rodriguez
Using Wikipedias Categories to Study the Coverage of a Users Interests for a Web Page
Recommender ................................................................................................................................. 7
Pei-Chia Chang and Luz M. Quiroga
Your Browser is Watching You: Dynamically Deducing User Gaze from Interaction Data ........... 10
David Hauger, Alexandros Paramythis, and Stephan Weibelzahl
Bridging Versioning and Adaptive Hypermedia in the Dynamic Web .......................................... 13
Evgeny Knutov, Mykola Pechenizkiy, and Paul De Bra
MyExperiences: Visualizing Evidence in an Open Learner Model ................................................ 16
Barbara Kump, Christin Seifert, Gnter Beham, Stefanie Lindstaedt, and Tobias Ley
A User Model to Predict Web Page Viewing Behavior ................................................................. 19
Rui Li, Evelyn Rozanski, and Anne Haake
Information Filtering Using Fragmented Preference Modeling for Individuals and Groups: A
Case Study ..................................................................................................................................... 22
S. Loeb and E. Panagos
An Adaptive Game to Train Users with Special Needs to Make Decisions: Using Public
Transportation .............................................................................................................................. 25
Estefana Martn, Silvia Calvo, and Rosa M. Carro
Towards Adapting Group Activities in Multitouch Tabletops ....................................................... 28
Estefana Martn and Pablo A. Haya
Enriching TV Experience: TV-viewing Support System Considering Both Individual and Family
Preferences ................................................................................................................................... 31
Chihiro Ono, Kazushi Ikeda, Gen Hattori, Kazunori Matsumoto, and Yasuhiro Takishima
Patient Condition Modeling in Remote Patient Management: Hospitalization Prediction ......... 34
Mykola Pechenizkiy, Aleksandra Tesanovic, Goran Manev, Ekaterina Vasilyeva,
Evgeny Knutov, Sicco Verwer, and Paul De Bra
Clustering of Web Users Using the Tensor Decomposed Models ................................................ 37
Rakesh Rawat, Richi Nayak, and Yuefeng Li
Towards Incorporating Confidence Measure into User Modeling ............................................... 40
Michael Yudelson, Peter Brusilovsky, Antonija Mitrovic, and Moffat Mathews
Demonstrations
Scrutinizing User Interest Models with IntrospectiveViews ......................................................... 43
Fedor Bakalov, Birgitta Knig-Ries, Andreas Nauerz, and Martin Welsch
Demonstration: Adaptive Agent-Based Learning of Intercultural Communication Skills ............ 46
W. Lewis Johnson and Alicia Sagae
GDR: A Language for Flexible Interlinking, Integrating, and Enriching User Data ........................ 49
Erwin Leonardi, Jan Hidders, Geert-Jan Houben, Fabian Abel, and Eelco Herder
A Motivational Mobile Application Toolset for Emotional Eating ................................................ 52
Kendra Markle and Lorin Wilde
Learning by Teaching SimStudent ................................................................................................. 55
Noboru Matsuda, Kenneth Koedinger, William Cohen, Victoria Keiser, Rohan Raizada, and
Gabriel Stylianides
MARY: A Personalised Virtual Health Trainer ............................................................................... 58
Hien Nguyen and Judith Masthoff
A Generic Privacy-Enhancing Personalization Infrastructure ....................................................... 61
Yang Wang and Alfred Kobsa
SocConnect: Intelligent Social Networks Aggregator ................................................................... 64
Yuan Wang, Jie Zhang, and Julita Vassileva
A Bayesian Hierarchical Preference Model for
Context-Aware Recommendations
Hideki Asoh
1
, Yoichi Motomura
2
, and Chihiro Ono
3
1
Intelligent Systems Research Institute, AIST
AIST Tsukuba Central 2, 1-1-1 Umezono, Tsukuba, Ibaraki 305-8568, Japan
h.asoh@aist.go.jp
2
Center for Service Research, AIST
AIST Tokyo Waterfront, 2-41-6 Aomi, Koto-ku, Tokyo 135-0064, Japan
y.motomura@aist.go.jp
3
KDDI R&D Laboratories, Inc., 2-1-15 Ohara, Fujimino, Saitama 365-8502, Japan
ono@kddilabs.jp
Abstract. A Bayesian hierarchical model is applied to the context-
aware collaborative recommendation problem. The proposed method treats
users, items, and contexts symmetrically, in contrast to existing context-
aware extensions of collaborative lters, which treat them asymmetri-
cally. Evaluation using internet questionnaire data demonstrates that the
proposed method outperforms conventional collaborative ltering (CF)
models and is very stable, even when the number of ratings is very small.
1 Introduction
Incorporating context-awareness of users preferences is important for recom-
mender systems[1]. In particular, recommender systems for mobile devices should
cope with changing contexts of users. We have constructed attribute-based pref-
erence models using Bayesian networks and developed a context-aware movie
recommender system for cell phone users using the model[2]. In this paper, we
propose to apply a simple Bayesian hierarchical model for context-aware collab-
orative ltering (CF).
There have been several studies extending CF to handle dierent contexts[1,
4]. Most of these reduce 3-way (user, item and context) rating tables into 2-way
tables and apply conventional CF methods. Here we treat 3-way rating tables
directly by modeling the preferences/ratings with a simple Bayesian hierarchical
model.
2 Bayesian Hierarchical Preference Models
Bayesian hierarchical modeling is a very eective tool for simultaneous estima-
tion of several parameters over similar units, and is used to capture heterogeneity
of subjects in areas such as marketing and ecology[5]. In the context of user mod-
eling, estimating personalized parameters
i
using observed responses r
ij
from
1
N users is a typical case for which the following model is eective:
r
ij
p(r|
i
)p(|)p(), i = 1, , N, j = 1, , K
i
There have been several applications of hierarchical modeling to information
ltering problems[6]. Most of them are hierarchical extensions of content-based
models. We propose to apply the following simple linear Gaussian hierarchical
model directly to the problem of collaborative prediction of ordinal ratings r
u,c,s
by user u for contents c in context s.
r
ucs
normal(
ucs
, 1/), gamma(, ),

ucs
=
0
+
u
+
c
+
s
,
0
normal(, ),

u
normal(0, 1/

),

gamma(, ),

c
normal(0, 1/

),

gamma(, ),

s
normal(0, 1/

),

gamma(, ).
3 Experimental Result
We applied the proposed model to our food preference dataset collected by means
of an internet questionnaire in December 2008[3]. In the dataset, 212 subjects
rated 20 food menus in 3 dierent conditions (hungry, normal and full) using
a 5-point scale. Each subject rated 5 menus in 2 conditions. Hence, the total
number of ratings was 2120. We divided the dataset into training data and test
data as follows: First, we chose 50 test users randomly; then, we randomly left
some ratings out of the 10 ratings of each test user for testing. The rest of the
ratings are used as training data. We change the number of ratings left out per
test user from 1 to 9 as a control parameter in the experiment.
We repeated the data division 10 times and evaluated the mean of the mean
absolute error (MAE) and mean squared error (MSE) of the rating prediction
for the left out test data. For comparison, the performances of conventional
correlation-based user-based collaborative ltering (CF), and the hierarchical
model without the
s
term (context-unaware) (HB2) are also shown. In the CF,
we treat items in dierent contexts as dierent items. We use WinBUGS to make
Markov Chain Monte Carlo inferences. Figure 1 summarizes the results. We set
= 2.0, = 10, = 2, = 1, however, the results are robust with respect to
these parameters.
The results demonstrate that the proposed simple hierarchical model out-
performs conventional CF. In particular, when the number of omitted ratings
is large, the improvements of the proposed model become larger. This implies
that the hierarchical model is very eective at coping with the cold start prob-
lem. Improvements of the MSE when considering context were also shown. We
applied paired t-test and conrmed that the MSE of the proposed method is
smaller than that of CF and HB2 at the 5% level even in the case that the
number of the left out sample is 1.
2

Fig. 1. MAE and MSE of the predicted ratings for test data
4 Conclusion and Future Work
We proposed to apply a simple Bayesian hierarchical model to a context-aware
collaborative ltering problem and show the eectiveness of the model using
internet questionnaire data regarding food preferences. Although this analysis is
preliminary, the results are promising.
Future work will include integrating collaborative and content-based ltering
within the proposed framework, seeking faster algorithms applicable to large-
scale datasets and applying non-Gaussian models for ordinal categorical data.
Acknowledgments. We thank Dr. Shigeyuki Akiba, President and CEO of
KDDI R&D Laboratories Inc., for his continuous support of this study. This
work was supported in part by JSPS KAKANHI 20650030.
References
1. Adomavicius, G., Sankaranarayanan, R., Sen, S., and Tuzhilin, A.: Incorporating
contexual information in recommender systems using a multidimensional approach.
ACM Transactions on Information Systems 23(1), 103145 (2005)
2. Ono, C., Kurokawa, M., Motomura, Y., and Asoh, H.: A context-aware movie
preference model using a Bayesian network for recommendation and promotion.
In: Proceedings of UM 2007, LNCS vol. 4511, pp. 247257, Springer-Verlag (2007)
3. Ono, C., Takishima, Y., Motomura, Y., and Asoh, H.: Context-aware preference
model based on a study of dierence between real and supposed situation data. In:
Proceedings of UMAP2009, LNCS vol. 5535, pp.102113 (2009)
4. Baltrunas, L. and Ricci, F.: Context-based splitting of item ratings in collaborative
ltering. In: Proceedings of the 3rd ACM Conference on Recommender Systems,
pp. 245248 (2009)
5. Congdon, P.: Bayesian Statistical Modelling, Second Edition, Wiley (2006)
6. Yu, K., Schwaighofer, A., Tresp, V., Ma, W.-Y., and Zhang, H. J.: Collaborative
ensemble learning: combining collaborative and content-based information ltering
via hierarchical Bayes. In : Proceedings of the 19th Conference on Uncertainty in
Articial Intelligence (2003)
3
WAI-WEB: Traveler assistance adapting user interface
and routes to people with cognitive disabilities
Linda Barros, Luis Rojas, Pablo A. Haya, Rosa M. Carro, Pilar Rodriguez


Escuela Politcnica Superior, Universidad Autnoma de Madrid
C/ Francisco Toms y Valiente 11
28049 Madrid, Spain
{lindaelizeth.barros, luisalberto.rojas}@estudiante.uam.es,
{pablo.haya, rosa.carro, pilar.rodriguez}@uam.es
Abstract. This paper presents an ongoing work named WAI-Web: a traveler
assistance adaptive system focused on users with cognitive disabilities. Users
have different profiles that reflect their cognitive skills and route preferences. In
addition, user interface is adapted by varying the application content and
dialogue. Routes are also personalized according users profiles. Finally, a
caregiver configuration tool supports the users profile creation.
Keywords: User modeling, Adaptive User Interface, Cognitive impairments,
Public Transportation Service
1 Introduction
One of the most common issues for people with cognitive disabilities is facing the
difficulty of traveling through the city. Public transport networks are a very complex
large-scale system. For a person with cognitive disabilities, learning a new route in
those scenarios can take several months [1]. Besides, maps and schedules are the main
tools for understanding those kind of networks, but they are not designed bearing in
mind the cognitive skills of this collective. In those standard maps and schedules, the
information is presented globally and using a high level of abstraction. However,
users with cognitive disability prefer sequential representation and more concrete
indications. Additionally, there are extra considerations that are also important such as
the difficulty of taking decisions when something unusual happens, i.e., the line that is
used for going to work is broken down and it is necessary to calculate an alternative
route.
Thus, cognitive assistance during traveling is required. New tools have been
designed for facilitating this task such as Mobility Agent [2], Opportunity Knocks [3],
Indoor Wayfinding [4], or WADER [5]. This paper presents a complementary
approach: WAI-Web. It is an ongoing work that focused on adapting both the
calculated route and the user interface. WAI-Web is based on WAI-Routes [6]. The
main goal of WAI-Web is to calculate the best route between two coordinates in
space using the public transport network. The system supports a variety of filters that
4
can be used for personalizing the output route. Possible examples are to eliminate a
certain station (I want to go from P1 to P2 without going through P3 subway station,
since it is a complex transfer) or to avoid a certain line (I want to go from P1 to P2
trying to avoid the L1 subway line since it is usually crowed).
2 Adapting routes and user interface
WAI-Web is focused on people with cognitive disabilities. In particular, we are
working with students of the Universidad Autnoma de Madrid in the first and second
course of an employability project for cognitively impaired population. These are
young people with intellectual disability for different reasons: Down's syndrome
(primary trisomy and mosaicism), Turner's syndrome, cerebral paralysis and
encefalophaties with not specified aetiology. Moreover, the system can also benefit to
occasional travelers and elderly people.
It is assumed that each user has associated one or more caregivers that are in
charge of creating and configuring the different profiles and groups the user can
belong to.
There are three different entities in the application model: users, groups and
favorite destinations. Each user has a particular profile that defines constrains for
calculating the routes and the presentation options. A user can belong to several
groups, and each users group shares a common profile that complements the personal
one. Finally, each user can have several predefined favorite destinations.
Regarding to the cognitive disability, users are classified in three different levels:
low, moderate and high. These levels have been identified by interviewing several
experts on disabilities. Each users level is associated to particular cognitive needs,
and determines the user interface appearance and dialogue.
WAI-Web allows configuring three different kinds of profile for each user. Each
profile is associated to one of the aforementioned entities and it is composed by a set
of constraints where each constraint corresponds to one of WAI-Routes filters [6]. For
instance, a constraint can be Avoid to take line 6, or Do not transfer in station A or
B. Besides, constraints can be modified by temporal conditions such as Do not use
tram during the weekend.
Thus, each user can have a particular profile, several group profiles and one
different profile for any of her favorite destinations. Route calculation is adapted to
each user according to users profiles. The system combines the constraints belonging
to different profiles according to the following rule: If the route destination coincides
with one of the favorite destinations, only the corresponding constraints are applied.
Otherwise, the aggregation of personal and group constraints is applied.
User interface is adapted to suit both the content presentation and the application
dialogue according to the users cognitive skills. Thus, the final layout has different
degrees of verbalization varying from iconic to textual representation. Start and
destination point selection is presented in an either global (all the information in one
screen) or sequential mode (one-by-one selection). And, finally, routes are showed in
three different ways: linear, alone or real map. The first one presents the route stops in
a horizontal or vertical line (depending on the user preferences). The second one
5
shows an alone and simplified version of the route but similar to the original map.
Lastly, the third one highlights the route in the real map.
At present, we have already developed a first prototype composed by two modules:
Adaptation engine and Caregivers Configuration Tool. Firstly, the adaptation engine
contains the user model. It is responsible of: a) choosing the right presentation and
dialogue settings for each user, b) receiving the start and destination points and,
tailoring the query depending on the users profile. Secondly, the Caregivers
Configuration Tool allows the caregiver to define the user model including users
profiles (personal, group and favorite destination), and cognitive level. The final
result can be accessed by standard web browsers.
4 Conclusions and future work
This paper presents a traveler assistance adapted to users with cognitive disabilities.
User interface is adapted according to the users cognitive skills. The route calculation
is also tailored, presenting different output routes depending on users profiles. As
future work, user activity on the system will be monitored, so if cognitive skill
improvements are detected, the cognitive level can be upgraded.

Acknowledgments. This work has been partially funded by the following projects:
HADA (Ministerio de Ciencia y Educacin de Espaa, TIN2007-64718), Vesta
(Ministerio de Industria, Turismo y Comercio de Espaa, TSI-020100-2009-828) and
eMadrid (Comunidad de Madrid, S2009/TIC-1650).
References
1. Carmien, S., Dawe, M., Fischer, G., Gorman, A., Kintsch, A. and Sullivan, J. F.: Socio-
technical environments supporting people with cognitive disabilities using public
transportation. ACM Trans. on Computer-Human Interaction, 12(2), pp. 233--262 (2005)
2. Repenning, A. and Ioannidou, A.: Mobility agents: guiding and tracking public
transportation users. In Proceedings of the Working Conference on Advanced Visual
interfaces AVI 06, pp. 127--134. ACM, New York, NY (2006)
3. Patterson, D. J., Liao, L., Gajos, K., Collier, M., Livic, N., Olson, K., Wang, S., Fox, D.
and Kautz, H: Opportunity Knocks: A System to Provide Cognitive Assistance with
Transportation Services. UbiComp 2004: Ubiquitous Computing, pp. 433--450 (2004)
4. Liu, A., Hile, H., Kautz, H., Borriello, G., Brown, P.A., Harniss, M. and Johnson, K.:
Indoor Wayfinding: Developing a Functional Interface for Individuals with Cognitive
Impairments. In Computers & Accessibility, ASSETS 2006, pp. 95--102 (2006)
5. Tsai, S.: WADER: a novel wayfinding system with deviation recovery for individuals with
cognitive impairments. In the 9th international ACM SIGACCESS Conference on
Computers and Accessibility, ASSETS 2007, pp. 267--268. ACM, New York, NY (2007)
6. Angls-Alcazar, L., Haya, P.A. and Carro, R.M.: WAI-ROUTES: A route-estimator
system for aiming public transportation users with cognitive impairments. In the Sixth
IASTED International Conference Advances in Computer Science and Engineering
(ACSE 2010), pp. 167--174 (2010)
6
Using Wikipedias Categories to Study the Coverage of
a Users Interests for a Web Page Recommender
Pei-Chia Chang, and Luz M. Quiroga
Information & Computer Sciences, University of Hawaii at Manoa,
1680 East-West Road Honolulu, HI 96822, USA
{pcchang, lquiroga}@hawaii.edu
Abstract. One major concern in the acquisition of user models is its ability to
elicit user interests. A users interests may be multiple and diverse, or few and
focused within a domain. We suggest adding an interest coverage dimension
for semantic user modeling. The Diversity Index is proposed to represent the
diversity of a users possible interest areas based on the selected categories
from Wikipedia. This paper evaluates the index as an instrument to identify and
represent a users interest coverage.
Keywords: user modeling, personalization, diversity, Wikipedia, recommender
1 Introduction and Related Work
In this paper, we are interested in using Wikipedias categorical structure to study the
coverage of a users interests to then allow web page recommendations across web-
sites. We define interest coverage as the semantic association among identified cate-
gories of interest to a user within a selected domain. Wikipedias content is highly
accessible and constantly changing; therefore, we expect using its categories as a
shared knowledge source and protocol would be a more dynamic and flexible plat-
form for user modeling.
The importance of interest coverage is dual. On one hand, the coverage of a users
interests may imply the users expertise in a domain. Page recommendations may
consider the expertise and avoid recommending an introductory page to an expert
user even if the page is semantically relevant. On the other hand, through a recom-
mender system that considers diversity, we intend to increase the possibility of infor-
mation encountering and discovery [3]. Information encounter is broadly defined as
the chance or memorable experience to gain valuable or interesting information out of
a random exploration.
Related studies on promoting diversity in recommendations include Symths discus-
sion between the tradeoff of similarity and diversity [4] and Zieglers effort in look-
ing into intra similarity among different recommendation lists [6]. Both studies are
concerned with the diversity given a selected topic, which differs from our focus on
7
the diversity of a users interests in different topics within a domain. The identifica-
tion of a users interests is as important as recommendations based on content similar-
ity. Another similar attempt to elicit user interests is found in Szomszor et als work
[5]. They adapt Wikipedias categories to consolidate user profiles, particularly tags,
from two social networking sites: del.icio.us and Flickr. One major distinction is that
our work takes the content-based approach while folksonomy analysis is Szomszor et
als primary effort. Content-based approach applies to any website, which does not
limit us to social websites that rely on collaborative filtering or tagging. Furthermore,
our work emphasizes examining the coverage of a users interests while interest ac-
quisition is Szomszor et als focus.
2 Method Description and Preliminary Result Discussion
Using the categorical vector-based model we derived from Wikipedia in the computer
science (CS) domain [2] we model a users interests via associating web usage pages
from various websites with the selected categories from Wikipedia. We classify user
interests into Wikipedias categories. The classification captures the closest mapping
and identifies the most promising categories of interest to the user. We further make
use of the topological relation of categories and construct a minimal spanning tree of
such promising categories. A variable Diversity Index (DI) is defined as the cumula-
tive weights of all edges in the tree. Edges are weighed differently according to the
relation (hierarchical, adjacent, or associative) between two categories. In addition to
the index, clusters of promising categories are identified as well. DI is used to priorit-
ize recommended pages that are selected by the cosine similarity measurement. This
measurement commonly applies to content based recommendations. Our system al-
ternates recommendations from one promising category to another or other adjacent
categories at a frequency rate proportional to the users DI value. Thus, a user with a
higher DI value receives recommendations in more categories or topics.
Table 1. Diversity and Interest Clusters
# Topic(s) DI Interest Clusters
1 Databases(DB) 41 DB9, CS organization3
2 HCI 35 HCI5, Graphics3, SE3, Computer architecture(CA)3
3 Algorithm 35 Algorithm13, Theoretical CS3
4 Software(SE) 77 SE5, DB3, HCI3, CA3
5 Data Mining(DM) 65 Algorithm8, DB4, Artificial intelligence(AI)3
6 SE\HCI 70 SE5, HCI4, DB3, CA3
7 SE\Algorithm 64 Algorithm9, SE4, Theoretical CS3
8 SE\DM 67 SE4, DB4, CA3, AI3, Algorithm3
9 Participant 1 65 DB6, Computer security4, Computer programming3
10 Participant 2 84 DB6
Regarding evaluation, we examine DIs accuracy by controlling the input values, i.e.
a users topical interests, and then validate the output values. We anticipate a low DI
8
value for focused users and a high DI value for diverse users. We created 8 simu-
lated users by respectively selecting 10 pages [1] to represent each users interests in
computer science (CS) related topics, as listed in Table 1. Additionally, we had 2 CS
participants who provided 20 usage pages. Table 1 displays the DI values and the
clusters with cluster size underlined. Larger cluster sizes imply stronger concentration.
According to Table 1, not only does the DI approximately quantify the coverage of a
users topical interests, but topical clusters that the user focuses on are identified also.,
both of which help to comprehend the semantics of users. Simulated users (#1-3) with
a focused topic have lower DI values than users with multiple topics (#6-8). Note that
DI values depend on the nature of topics as well. Software and data mining (#4-5) are
application dependent and diverse, which thus have higher DI values. Both partici-
pants (#9-10) are interested in topics of the listed cluster(s). Their DI values indicate
their interests in multiple topics, which were confirmed by them. Participant 2 has a
work-related concentration in the DB area in addition to other interests.
Diversity Index is an indicator for the coverage of a users interests within a domain
but not an absolute value. However, the interest coverage dimension has implications
for web page recommendations. For users who have diverse interests within a domain,
they could receive recommendations in more topics. Conversely, focused users
could receive recommendations in the advanced area of an identified topic. Both
cases increase the chance of discovering novel or serendipitous information. Other
potential applications include formulating a personal ontology and facilitating exper-
tise discovery among users, among other things.
References
[1] Complete page list: http://www2.hawaii.edu/~pcchang/ICS699/umap.html
[2] Chang, P. C., & Quiroga, L. M. An Ontological User Modeling Approach Using Wikipe-
dia's Content for Cross-System Personalization. In Proceedings of the IUI2010
Workshop on User Data Interoperability in the Social Web, (2010) 30-33
[3] Erdelez, S. Information Encountering: A Conceptual Framework for Accidental Informa-
tion Discovery. In Proceedings of the International Conference on Research in In-
formation Needs, Seeking and Use in Different Contexts (1997) 412-421
[4] Smyth, B., & McClave, P. Similarity Vs. Diversity. Paper presented at the Proceedings of
the 4th International Conference on Case-Based Reasoning: Case-Based Reasoning
Research and Development. (2001)
[5] Szomszor, M., Alani, H., Cantador, I., O'Hara, K., & Shadbolt, N. Semantic Modelling of
User Interests Based on Cross-Folksonomy Analysis. Paper presented at the Proceed-
ings of the 7th International Conference on The Semantic Web. (2008)
[6] Ziegler, C.-N., Lausen, G., & Schmidt-Thieme, L. Taxonomy-Driven Computation of
Product Recommendations. In Proceedings of the ACM international conference on
Information and knowledge management, (2004) 406-415
9
Your Browser is Watching You: Dynamically Deducing
User Gaze from Interaction Data
David Hauger
1
, Alexandros Paramythis
1
, and Stephan Weibelzahl
2
1
Institute for Information Processing and Microprocessor Technology
Johannes Kepler University, Altenbergerstrasse 69, A-4040, Linz, Austria
{hauger, alpar}@fim.uni-linz.ac.at
2
National College of Ireland, Dublin, Ireland
School of Computing, National College of Ireland, Mayor Street, IFSC, Dublin 1, Ireland
sweibelzahl@ncirl.ie
Abstract. Most user models in adaptive hypermedia systems rely on server-side
data collection, such as pages visited or links selected. The work reported in
here seeks to expand the inferences that can be made on the client side, by ob-
serving and analyzing patterns in the users interactive behavior such as mouse
moves, clicks, text selection and scrolling. These patterns can be used to esti-
mate with high levels of accuracy what the user is reading on a page. In an em-
pirical study involving gaze tracking of 13 participants we validated these pat-
terns and analyzed different prediction models. We established that a high fre-
quency of mouse moves on a page, clicks on text and the mouse being in mo-
tion strongly contribute to the accuracy of predictions of what users have read.
Key words. interaction monitoring, gaze modeling, eye-tracking, empirical
study, mouse activity
1 Interaction Data in Adaptive Hypermedia
Traditionally adaptive hypermedia systems (AHS) use server-side data collection
for inferences about what the users have read, are interested in, etc. The users client-
side behavior has long been identified as a potential additional source of information
towards the goal of increasing the accuracy of assumptions regarding what they really
read while at a page. Previous work exploited client-side data collection to identify a
general level of activity [1], the reading time [2], or learning types [3] based on
patterns in the mouse behavior. Rather than making inferences on pages as a whole, it
is possible to treat fragments of a page separately [4]. We decided to go one step fur-
ther, and examine whether it is possible to increase the accuracy of predicting what a
user was looking at while at a page, by identifying and interpreting specific patterns in
the users interactive behavior.
We introduced a number of hypotheses on usage patterns users might show and,
based on their observation, what additional inferences could be made on eye gaze.
In order to test our hypotheses, we designed a study that would allow us to com-
pare users reading behavior when encountering different types of text, with their in-
10
teractive behavior while reading these texts through a browser. Reading behavior was
determined through eye-tracking, whereas interactive behavior was recorded through
the purposely developed JavaScript library. Participants were given four tasks to per-
form, each based on a different type of text typically encountered online (instructions
and additional information for a board game, a set of search results, a health-related
article, and a set of news items). All material was presented through a TFT screen,
running a 1280x1024 resolution. The preliminary results presented below are based
on 13 participants (6 male, 7 female) completing the first task.
The study itself was organized as follows: After answering a questionnaire on de-
mographic data and prior knowledge, participants completed a reading speed test.
Next, participants were instructed to read a series of four different types of text on the
screen. The first text, and the one addressed in this paper, introduced participants to
the Game of Go. The text was split into seven pages. It comprised text (ca. 7010
words), graphics (11) and pictures (5).
While participants were reading, two types of data were recorded. Firstly, the Ja-
vaScript library, embedded in each web page, recorded all mouse and keyboard beha-
vior, including the mouse position, mouse clicks, keys pressed and scrolling events.
Secondly, using the eye-tracker, we recorded the actual gaze position.
2 Results
In total 112 page requests have been observed with a page being visited for 2 to 1096
seconds with a mean of 122 (=116s). On average each user spent 17.5 minutes on
the information on the game of Go. For this study we have defined a mouse move
to be any set of changes in the mouse pointers position, preceded and followed by at
least one second of idle time (i.e., time during which the mouse pointers position
does not change).
Even without looking at particular interaction patterns, the current mouse position
can give a very rough indication of where the user is looking: mouse and eye position
are correlated, both horizontally (r=.101, N=89739) and vertically (r=.250, N=89739).
The vertical correlations being much higher than the horizontal correlation might be
due to users not typically following their gaze with the mouse from left to right when
reading, but pointing at the paragraph.
The first step towards improving upon this baseline correlation was the insight that
some people use the mouse a lot, while others park it as long as they do not need it.
On pages where the mouse is moved frequently it should be easier to estimate where
people are looking. For pages where users moved their mouse frequently, the mouse
pointers position is strongly correlated with the position of the users gaze. The fre-
quency ratio (FR) denoting the proportion of time moving the mouse on a page was
used as a filter for selecting pages. We applied filter settings between 25% and 75%.
Is it possible to improve prediction of gaze by considering the frequency of mouse
usage on a page? The higher the percentage of mouse movements, the lower the dis-
tance between mouse and gaze positions (r=-.299; N=89739). When events are fil-
tered by the level of frequency of mouse movements, the correlation increases even
further (see Table 1). The more restrictive the filter, the higher the correlation. In ac-
11
cordance with the baseline, the correlations in the vertical direction are higher than in
the horizontal. The correlations get even higher when limiting the analysis to events
when the mouse is actually in motion, both in horizontal (e.g., r
>25%
=.746, N= 36202
and r
>75%
=.777, N= 21270) as well as in vertical (e.g., r
>25%
=.521, N= 36202 and
r
>75%
=.580, N= 21270) direction. Further to the above, results also showed that some
events like clicks on non-link text and text selections are very well suited to identify
whether the related paragraph has been read, although the improvement of prediction
they afford comes at the expense of limited coverage.
In summary, we found that certain patterns in users interactive behavior can be
used to approximate their gaze position. Based on these results we are currently de-
veloping a prediction algorithm that is used by the client-side monitoring mechanism
to deduce what parts of pages users have focused on. The algorithm uses weighted
interaction pattern-based indicators that combine several of the reported results. We
plan to implement it as a JavaScript library that can then be integrated into any Adap-
tive Hypermedia System to improve the granularity and accuracy of user models with
respect to what users have really looked at while on a page.
Table 1. Correlations of x and y positions of mouse cursor and eye gaze, depending on the
frequency of mouse usage. Regression model: eye pos. = (weight * mouse pos.) + constant
Frequency of
mouse moves
Correlation Regression model: mouse on eye
r
eye vs mouse
N constant weight sig.
vertical baseline .250 89739 345.133 0.228 .000*
frequency > 25% .608 39134 211.082 0.567 .000*
frequency > 50% .658 21906 173.328 0.613 .000*
frequency > 75% .746 16360 165.461 0.666 .000*
horizontal baseline .101 89739 577.499 0.075 .000*
frequency > 25% .393 39134 385.193 0.386 .000*
frequency > 50% .493 21906 241.912 0.604 .000*
frequency > 75% .560 16360 188.254 0.727 .000*
Acknowledgments. The work reported in this paper is partially funded by the Adap-
tive Support for Collaborative E-Learning" (ASCOLLA) project, supported by the
Austrian Science Fund (FWF; project number P20260-N15).
References
1. Goecks, J., Shavlik, J.W.: Learning Users Interests by Unobtrusively Observing Their
Normal Behavior. In: Int. Conf. on Intelligent User Interfaces Proceedings of the 5th Int.
Conf. on Intelligent User Interfaces. (2000) 129-132
2. Spada, D., Snchez-Montas, M.A., Paredes, P., Carro, R.M.: Towards Inferring Sequen-
tial-Global Dimension of Learning Styles from Mouse Movement Patterns. In: AH. (2008)
337-340
3. Cha, H.J., Kim, Y.S., Park, S.H., Yoon, T.B., Jung, Y.M., Lee, J.H.: Learning Styles Diag-
nosis based on User Interface Behaviors for the Customization of Learning Interfaces in an
Intelligent Tutoring System. In: Intelligent Tutoring Systems. Vol. 4053. (2006) 513-524
4. Hauger, D.: Using Asynchronous Client-Side User Monitoring to Enhance User Modeling
in Adaptive E-Learning Systems. In: UMAP 09: Adaptation and Personalization for Web
2.0, Trento, Italy (2009)
12
Bridging Versioning and Adaptive Hypermedia in the
Dynamic Web
Evgeny Knutov, Mykola Pechenizkiy, Paul De Bra
Eindhoven University of Technology, Department of Computer Science
PO Box 513, NL 5600 MB Eindhoven, the Netherlands
{e.knutov, m.pechenizkiy}@tue.nl, debra@win.tue.nl
Abstract. Web Dynamics has been recently considered in the context of the
analysis of how people search and re-search information on the web. There are
lots of challenges and opportunities when considering user behaviour. In this
paper we propose the way to tackle some of them by applying versioning
methodologies (as a backend solution) in the context of content changes, user
re-visitations and re-searches on the web, as well as Adaptive Hypermedia
(AH) techniques to overcome visualisation issues (as a frontend solution).
Essentially we bridge versioning and AH in the field of Web Dynamics
showing how versioning helps to make the adaptation scrutable.
Keywords: adaptive hypermedia, versioning, web dynamics, user modelling.
1 Introduction
In a dynamic information environment such as the Web the understanding of the
constantly evolving content is becoming an issue [1]. The main problem is that user
behaviour is influenced by the fact that changes to the content are observed. These
changes may cause re-visitation and re-searches for the same information.
In order to cope with the information overload issues, changes and evolving
structures on the web we may think of using versioning approaches and technologies.
Besides this, the rates of web re-visitations and re-searches (which are estimated to be
very high [1, 3], e.g. re-visitation rates are about 50-80%) incite us to investigate
versioning in the area of user visitation and search behaviour.
We essentially consider a number of scenarios of bridging versioning
methodologies with AH. The first one is the backend solution, which employs
conventional source control approaches. Secondly, we consider the front-end solution
that uses AH methods and techniques, presenting versioned information to the user.
And as a result we think that these two can be merged moving towards an intelligent
versioning cache system or a proxy that would serve as an Adaptive Hypermedia
System (AHS) and would function as a proxy between a conventional web
information system and a user (user profile) retrieving, dispatching and presenting
adapted information (web content, user queries, sites visited).
13
2 Versioning to Support AH in a Dynamic Web Environment
Considering the basic concepts of versioning we can come up with the following
classes of operations which will reflect typological and structural types of changes
and capture changes in the dynamic web environment. The following taxonomy of
changes was extended in the field of AH [4]. Here we extend and elaborate it in terms
of Web Dynamics, describing properties and potentials of versioning operations
which could be used mostly as a backend solution to store, keep track of changes and
retrieve them for further analysis: Transformation is a set of actual changes;
Conceptual changes refer to changes of the structure, relationships or presentation
aspects; Descriptive changes dealing with metadata describing the intentions, user
or author credentials, and reasoning behind the changes; Context changes describe
the environment in which the current update occurs and the environment where it is
valid.
Much stress has been placed on the visualization part of re-searches and re-visitation
cases, presenting mostly historical snippets in the search result list or browser
extensions which highlighted the difference in the newer versions of the document
[3]. We take a look at the visualization issues throughout AH research and use content
and presentation adaptation techniques introduced there [5] in order to support
versioning visualisation. At the same time one may consider different versions of
some content and apply the same techniques from the AH field taking into account
the fact that different versions just represent the different variants of the same
fragment as it is done in AHS.
3 Use-Case: User Web Activity in a Versioned Environment
UM.ver.1
user
Web page
New Web page
(ver.2.)
UM.ver.2
logging user bahaviour
UM.ver.N
Search
results 1
Search
results 2
Old version of Web
page (ver.1)
cached
Proxy: Retrieving and presenting
versioned results:
- history of queries with related
information
- difference between search results
comparing to the current query
- for the chosen result the difference
of web page is presented (between
cached and current version)
difference
difference
new
query
Page content
changed
Showing
versioned results
to the user
Fig. 1 User web activities in versioned environment
In Fig. 1 we present a use-case bridging versioning approaches used to track user
activities in the dynamic web environment with the AH presentation of the versioned
content (e.g. user search results history, changes in web pages, etc.).
14
We have a user, who searches and browses through the web. The initial state of UM
(UM.ver.1) starts accumulating search and visitation history. As the user continues
the interaction he posts a new query. An Adaptive Proxy retrieves the previous state
of UM and search queries and result lists with changes to UM. These changes are
processed and compared to be presented, providing an opportunity to see what other
related queries and results he has already done. As a result actual changes of a
particular page can be retrieved and presented using AH techniques. Search results
could also be presented using AH techniques and re-arranged according to a new
global or personalized ranking that may change over time for various reasons.
4 Innovative Aspects and Prospective Solutions
Here we would like to conclude and outline the advantageous, innovative aspects and
prospective versioning solutions in web dynamics. These are the following:
System Design and Authoring versioning helps to create, maintain, re-use,
reconcile concurrent versions of an application, model or a particular property or
value, saving authoring and design effort.
Storing versioning provides an efficient way to store changes, label and annotate
them. It facilitates convenient, hierarchical structure presentations and offers a
number of operations to handle it.
System Maintenance and Support structured changes and a number of operations
(e.g. merge, resolve, etc.) are sufficient to maintain and reconcile application
conflicts, inherit functionality between system versions.
Logging logging changes provides flexible playback possibilities and can serve as a
basis for system analysis. Logging user updates will provide a ground for user
behaviour comparison.
Analysis step-by-step system and user behaviour analysis is facilitated. Versioning
which provides hierarchical incremental logs exploits the transparency of the system
functionality and evolution of the user web environment.
References
1 Adar, E., Teevan, J., Dumais, S. T., Elsas, J. L.: The web changes everything: understanding
the dynamics of web content. In Proc. of 2nd ACM Int. Conf. on Web Search and Data
Mining (WSDM'09), pp. 282-291. ACM, New York, NY, (2009)
2 Teevan, J., Dumais, S. T., Liebling, D. J., Hughes, R. L.: Changing how people view
changes on the web. In Proc. of 22nd ACM Symposium on User interface Software and
Technology (UIST '09). pp. 237-246. ACM, New York, NY (2009)
3 Dumais, S. T., Cutrell, E., Cadiz, J. J., Jancke, G., Sarin, R., Robbins, D.: Stuff Ive Seen: A
system for personal information retrieval and re-use. In Proc. of SIGIR03, pp. 72-79 (2003)
4 Knutov, E., De Bra, P., Pechenizkiy, M., Versioning in Adaptive Hypermedia, In Proc. of
Workshop on Dynamic and Adaptive Hypertext: Generic Frameworks, Approaches and
Techniques, pp. 61-71. CEUR-WS.org, Aachen, (2009)
5 E. Knutov, P. De Bra and M. Pechenizkiy. AH 12 Years Later: a Comprehensive Survey of
Adaptive Hypermedia Methods and Techniques, New Review of Hypermedia and
Multimedia 15(1), pp. 5-38. Taylor & Francis, UK, (2009)
15
MyExperiences: Visualizing Evidence in an Open
Learner Model
Barbara Kump
1
, Christin Seifert
2
, Gnter Beham
2
, Stefanie Lindstaedt
1,2
,
Tobias Ley
2

1
Knowledge-Management Institute, TU Graz, Inffeldgasse 21a, 8010 Graz, Austria
2
Know Center, Inffeldgasse 21a, 8010 Graz, Austria
bkump@tugraz.at;
{cseifert, gbeham, slind, tley}@know-center.at
Abstract. When inferring a users knowledge state from naturally occurring
interactions in adaptive learning systems, one has to makes complex
assumptions that may be hard to understand for users. We suggest
MyExperiences, an open learner model designed for these specific
requirements. MyExperiences is based on some of the key design principles of
information visualization to help users understand the complex information in
the learner model. It further allows users to edit their learner models in order to
improve the accuracy of the information represented there.
Keywords: open learner model, information visualization
Introduction
Knowledge Indicating Events (KIEs) [1] have been suggested for diagnosing user
knowledge in an adaptive learning system. Similar approaches to assessing user
knowledge from different sources of evidence have been proposed (e.g., [2, 3]). KIEs
are naturally occurring actions from which evidence for a users knowledge level for a
concept is inferred. Because no additional interaction with the users is needed to
diagnose knowledge levels, the KIE approach is especially useful e.g., for work-
integrated learning. However, employing KIEs is based on assumptions which may
not be valid in all cases (e.g., a user who reads a text has learned about the
underlying concept). Moreover, the diagnosis of user knowledge based on KIEs is
not necessarily understandable for users.
Open learner models (OLMs) [4] enable users of adaptive learning systems to see
the details of the information held about them as well as the processes used to gather
this information. Thus, we regard OLMs as powerful means to enhance
understandability and accuracy of knowledge assessment based on KIEs. We present
MyExperiences, the OLM of the adaptive learning system APOSDLE. Because in
OLMs the way the information is displayed to the user is critical, the design of
MyExperiences is based on the principles of information visualization condensed in
Shneidermans well-known mantra Overview first, zoom and filter, then details-on-
demand [5].
16
MyExperiences: The Open Learner Model of APOSDLE
APOSDLEs open learner model, MyExperiences (Fig. 1), allows users to access their
learner model in order to understand how their knowledge level in certain concepts
was diagnosed, and to change its content, if necessary. Typically, ~100 concepts are
represented in the learner model. For each concept, one of three knowledge levels is
automatically diagnosed with KIEs: learner (e.g., asking for a learning hint), worker
(e.g., performing a task related to the concept), and supporter (e.g., being contacted).
In terms of data structures, the APOSDLE learner model is a forest of trees with
the knowledge levels being the roots of the trees. The children of each root
(knowledge level) are all concepts for which the corresponding knowledge level was
diagnosed. The children of each concept again are the KIEs that occurred for the
respective concept. Due to the large amount of data (number of knowledge levels,
concepts, and KIEs), we chose the tree map [6], a space-filling approach, as the core
technique. By combining the tree map (right part of Fig. 1) with the familiar tree view
(left part of Fig. 1) as coordinated multiple views, the user can get a quick overview
(tree map) and browse to details in a familiar way of interaction (tree view).


Fig. 1. The tree map-based open learner model of APOSDLE
This is how MyExperiences relates to the principles of information visualization
overview, zoom and filter, and details-on-demand: MyExperiences is divided into
three rows (with different colors), one for each knowledge level. Each row in
MyExperiences consists of the tree view (left) and the tree map view (right). Each tree
map gives an overview of all the concepts for the specific knowledge level. For a
concept, the number in brackets indicates the number of related evidences (KIEs).
The lighter the color of such a sub-rectangle, the higher is the relative occurrence of
this event for the concept. The aspect of filtering is realized with the search
17
functionality on the top left. The user has several possibilities to interact with
MyExperiences. Common interaction techniques like selection and zooming enable
the user to investigate the OLM, either as an overview or in detail. Zooming one step
into the tree map allows the user to understand why the specific knowledge level was
inferred for this concept. In this zoomed view, the user is provided with the additional
information about which evidences (KIEs) occurred how often for this specific
concept (details-on-demand). The user also can alter the knowledge level for any
concept and resets its own changes to the systems decisions.
Discussion and Future Work
Outcomes of informal interviews with APOSDLE users indicate that MyExperiences
improves understandability and accuracy of the APOSDLE learner model. Still, this
needs to be evaluated systematically in usability studies. Concerning the visualization,
we see two main avenues for follow-up. First, we plan to provide users with visual
clues about uncertainty in automatically diagnosed knowledge levels. Second, users
should be able to see the development of their knowledge levels over time. One
possible disadvantage of the current realization of MyExperiences is that concepts
without KIEs do not appear in the visualization and thus cannot be altered by the
users. This may be overcome in a future version.


Acknowledgements. The Know-Center is funded within the Austrian COMET
Program - Competence Centers for Excellent Technologies - under the auspices of the
Austrian Ministry of Transport, Innovation and Technology, the Austrian Ministry of
Economics and Labor and by the State of Styria. COMET is managed by the Austrian
Research Promotion Agency FFG. APOSDLE (www.aposdle.org) has been partially
funded under grant 027023 in the IST work programme of the European Community.
References
1. Lindstaedt, S. N., Beham, G., Kump, B., Ley, T.: Getting to Know Your User
Unobtrusive User Model Maintenance within Work-Integrated Learning Environments.
Proc. of ECTEL 2009, pp. 73-87, Springer, Berlin, Heidelberg (2009)
2. Kay, J., Kummerfeld, B. and Lauder, P.: 2002, Personis: a server for user models. In:
P.D. Bra, P. Brusilovsky, and R. Conejo (eds.): Proceedings of AH2002, pp. 203212.
3. Brusilovsky, P.: KnowledgeTree: A Distributed Architecture for Adaptive E-Learning.
WWW 04, pp. 104-113, ACM, New York, NY (2004)
4. Bull, S., Kay, J.: Student Models that Invite the Learner In: The SMILI Open
Learner Modeling Framework. Int. J. of AI in Education, 17, 2, 89-120 (2007)
5. Shneiderman, B.: The Eyes Have It: A Task by Data Type Taxonomy for Information
Visualizations. In: Proc. IEEE Symp. Visual Languages, pp. 336-343, IEEE Computer
Society Press, Washington (1996)
6. Shneiderman, B.: Tree visualization with tree-maps: 2-d space-filling approach. ACM
Transactions on Graphics, 11, 92-99 (1992)
18
A User Model to Predict Web Page Viewing Behavior
Rui Li, Evelyn Rozanski, and Anne Haake
Computer Human Interaction and Performance Support Laboratory,
Rochester Institute of Technology, 14623 Rochester, U.S.A.
{rxl5604,evelyn.rozanski,anne.haake}@rit.com
Abstract. When people explore Web pages they make extensive use of knowl-
edge about how and where the information tends to appear. This paper proposes a
probabilistic approach integrating both goal-directed and stimulus-driven mech-
anisms to predict users visual attention deployments during informational tasks.
Our model incorporates Web users expectations about information allocation
within Web pages into the model. An eye tracking experiment shows that our
model performs as well as or better than existing models.
Key words: Eye tracking, Probabilistic Modeling, Web design.
1 Introduction
Visual attention working as a selective lter constrains people to engage in detailed
inspection of a very limited spatial region [3]. People therefore have to view particular
regions in a serial manner. In this sense, an understanding of how users view Web pages
when they are seeking information has become an important research area in terms of
improving the efciency of the interaction between the human and the Web.
Fundamental knowledge about viewing behaviors on Web pages has been pro-
vided by previous empirical studies [1,5,6]. Based on these, viewing behaviors on Web
pages have been explored by various computational modeling approaches[7,8]. How-
ever, none of these approaches can integrate both the top-down and bottom-up cognitive
factors which are crucial in terms of visual attention deployments [2].
2 Probabilistic Model
An ideal observer will xate on the points within a Web page that have the highest
probability of containing the information to meet the goal which we call a target, given
the available Web information. Whether the information conveyed in a location is se-
lected as the target depends in part on its utility. This is a binary decision: either the user
nds the information provided at the region useful enough to xate on it over and over
again or not. In this sense, we introduce the latent variable U denoting utility. Thus, se-
mantic features and visual features are conditional independent given utility. The joint
probability distribution factorizes as follows:
p(G, U, V, S, L) = p(G|U, L)p(U|S, V )p(S)p(V )p(L) (1)
19
Given the observed Web page point i, we can represent the evaluation process as
p(G = 1, U = u
i
|V = v
i
, S = s
i
, L = l
i
), and propose that this evaluation process
as the users Selection Sensitivity SS. Following Equation 1 and Bayes rule, we can
divide the Selection Sensitivity probability function into a set of components that can
be interpreted in terms of different mechanisms involved in the information-seeking
behavior:
ss
i
= p(G = 1, u
i
|v
i
, s
i
, l
i
) (2)
=
p(u
i
|G = 1)
p(u
i
)
p(u
i
|v
i
, s
i
)p(G = 1|l
i
) (3)
The probability p(u
i
|G = 1) is deterministic (and equal to 1) because once point i
is considered to contain the target piece of information, its utility is determined. Then
the rst term 1/p(u
i
) is the self-information of point i. The second term p(u
i
|v
i
, s
i
) rep-
resents the constraint of users utility evaluation. The third term p(G|l
i
) is independent
of visual and semantic features and describes the users expectation about the possible
location of the target information within the Web page.
3 Experiment
Ten college students were recruited. Eye tracking was performed with the Applied Sci-
ence Laboratories (ASL) 6000 remote eye tracking system. 50 Web pages were dis-
played in random order and the participants were asked to search a piece of information
in each page. The system calibration was checked after every 10 Web pages. A xation
was detected by the software after steadily gazing in an area with a radius of 50 pixels
for at least 100 ms.
4 Results and Discussion
We adopted the performance evaluation strategy in [4]. Thus, for the evaluation of each
model, we used 100 random permutations of the prediction maps to calculate the area
under the ROC curves for Ittis visual saliency model, LSA-SF model and our proba-
bilistic model, and the values are 0.58, 0.62 and 0.67 respectively. Figure 1 displays two
representative cases with xation data map and all the model predictions.
Our model predicts that participants tend to read the organizational text which is an
optimal behavior as what an ideal observer should do (see Figure 1) in terms of Web
page viewing. This suggests the importance of navigational cues such as header bars
and organizational texts that help facilitate users seeking target pieces of information
on Web pages. From the results, we can also see that participants xation data are
often clustered in the center of Web pages, which was not accounted for by any of the
previous models.
5 Conclusion
We derived a probabilistic model in which the three critical cognitive processes were
combined to make better prediction on users visual attention deployment. In future
20
analysis, we will extend our model into a non-parametric analysis in order to investigate
the viewing behavior differences between individuals or types of Web pages.
(a) Web page (b) Fixation map (c) Bayesian (d) LSA-SF (e) Itti et al.
Fig. 1: Examples of predictions for qualitative comparison. From left to right, each row
contains: (a) An original Web page; (b) the xation map from observed data of 10
participants; (c) the Selection Sensitivity map generated by our probabilistic model; (d)
the semantic elds map produced by the LSA-SF model; (e) the saliency map from Itti
et al. work as a comparison.
References
1. Lorigo, L., Pan, B., Hembrooke, H., Joachims, T., Granka, L., Gay, G.: The Inuence of
Task and Gender on Search and Evaluation Behavior Using Google. Info. Processing and
Management: an Intl Journal. 142, 11231131 (2006)
2. Itti, L., Koch, C.: Computational Modelling of Visual Attention. Nature Reviews, Neuro-
science. 2, 194203 (2001)
3. Rayner, K.: Eye Movements in Reading and Information Processing: 20 Yeas of Research.
Psychological Bulletin. 124, 372422 (1998)
4. Zhang, L., Tong, H.M., Marks, T.K., Shan, H., Cottrell, G.W.: SUN: A Bayesian Framework
for Saliency Using Natural Statistics. J. of Vision. 8, 120 (2008)
5. Pan, B., Hembrooke, H.A., Gay, G.K., Granka, L.A., Feusner, M.K., Newman, J.K.: The De-
terminants of Web Page Viewing Behavior: An Eye-Tracking Study. In: Proc. of 2004 sympo-
sium on Eye tracking research and applications, pp. 147154. ACM, New York (2004)
6. Cutrell, E., Guan, Z.: What Are You Looking For? An Eye-Tracking Study of Information
Usage in Web Search. In: Proc. CHI 2007, pp. 407416. ACM, New York (2007)
7. Buscher, G., Cutrell, E., Morris, M.R.: What Do You See When Youre Surng? Using Eye
Tracking to Predict Salient Regions of Web Pages. In: Proc. CHI 2009, pp. 2130. ACM, New
York (2009)
8. Stone, B.P., Dennis, S.: Using LSA Semantic Fields to Predict Eye Movement on Web Pages.
In: Proceedings of the Twenty Ninth Conference of the Cognitive Science Society, pp. 665
670. Cognitive Science Society, Austin (2007)
21
Information Filtering Using Fragmented Preference
Modeling for Individuals and Groups: A Case Study
S. Loeb and E. Panagos
Applied Research, Telcordia
One Telcordia Drive, Piscataway, NJ 08854
{shoshi, thimios}@research.telcordia.com

Abstract
We present a preliminary study of information filtering that is focused on the
modeling of user profiles as collections of fragments that are time and context
sensitive and updated independently. In each situation the appropriate fragment(s) of
a user profile is activated, used, and updated. Furthermore, each profile fragment can
be derived from a different source and by a different method, such as explicitly stated
individual preferences or collaborative behavioral groupings.
Case Study
Today, the vast amounts of information available to individuals increase the need for
automated ways to select and deliver what users may consider to be relevant.
Researchers have been developing models that capture individual preferences, and
attention has been paid to the effect of time and context as well as the influence of
other individuals on information preferences and needs. With the explosion of
advanced services available on smart mobile devices, information filtering and
personalization can now be offered to individuals on the go with special emphasis on
relevancy based on location, time, and other contextual information (e.g., [1], [5]).

It has been recognized that more than one method for modeling users may be
needed in order to increase the accuracy of information recommendation. Magitti [1],
for example, offers both collaborative and contextual filtering to mobile users.
Contextual filtering is driven by sophisticated user modeling that combines various
direct and inferred user preference data, calendar, activity and other contextual
information. A hybrid model with, possibly, different weight given to the various
modeling methods showed to generate the most diverse and novel recommendations,
and it was proposed that, perhaps, users can provide input into what weight to give to
what modeling method. Additional works indicating the importance of mood, time
and context (e.g., [3] [4] [7]) also point to the mixed model approach where the
individual user information needs are not possess or determined by a single
22
monolithic model, but rather are constructed from fragments of models, each useful in
its own time, context and personal parameter range.
In this poster paper, we present the results of a preliminary study that put under the
microscope a group of 12 university students. The group met each week for eight
weeks to discuss different aspects of mobile services, including location-based
services, context-aware services, mobile social networking, and automotive
telematics. The students were asked to bring to each meeting an article of interest
from a source of their choice and discuss it for a brief time. Then in the following
meeting, they were given the complete list of articles selected by the group, and after
a brief review, they were asked to rank the articles on a scale from 1-10 based on
how interesting each article was to them. The intention of the study was to uncover
the various forces at play when an individual, who is also part of a group, selects
information of interest to her, and then ranks her level of interest in the selected
information in the context of information selected by the rest of the group.
Preliminary Results
In this section, we present our initial set of observations based on the information
collected during the 8 week study mentioned above. We were interested to see who of
these individuals (if any) picked articles the rest of the group consistently liked, who
consistently ended up not liking the articles they selected, how diverse were the
information sources from which the articles were selected, and how diverse were the
set of article subjects over time. In doing so, our objective was to gain understanding
of the effects of time, context, group membership, and availability of information on
individual preferences.
We were interested in the concept of fragmented user profile as mentioned above,
in the contribution of the collaborative, or cross user effects to improve the quality of
the delivered information and in the evolution of user profile over time. We analyzed
the data collected over the study period and studied the views of the 12 individuals on
both their own selections and the selections of other group members (Fig. 1- left). We
furthermore correlated and trended the emerging profile information and noticed the
cross user effects (Fig. 1 - right).

Fig. 1. Left - percentage of times individuals ranked their selection at the top or
among the top 3. Right - influence of selected topic. Highlighted cells contain highest
normalized ranking for each topic (lower score corresponds to higher ranking).
23

Conclusions
Being able to identify information that may be of interest to a user and meet her needs
is a challenging information filtering problem that hinges on user profile modeling. In
this work we presented a preliminary study involving 12 individuals over an 8 week
period, with each individual selecting a weekly news article in a specific domain from
information sources of their choice. We studied their views on their own selection as
well as selections of other group members and correlated the emerging profile
information. We noticed collaborative filtering and other cross user effects and
examined the shift in preferences over time and some of the preferences convergence
and grouping of the individual profiles.
Currently, we are expanding the study both in time and scope and incorporate our
observations in the CUSP context-aware prototype system we developed in the past
two years for the purpose of providing personalized recommendations related to user
activities, such as appointments and to-do tasks [2][6]. In particular, we are
incorporating fragmented and group profiles in CUSP by utilizing augmented-reality
techniques for the delivery of the filtered information.
References
1. Ducheneaut, N., Partridge, K., Huang, Q., Price, B., Roberts, M., Chi, E.,
Bellotti, V., Begole, B. Collaborative Filtering is Not Enough? Experiments
with a Mixed-Model Recommender for Leisure Activities. In G.-J. Houben et al.
(eds): UMAP 2009, LNCS 5535: UMAP 2009, 295-306 (2009).
2. Falchuk, B., Loeb, S., Panagos, E. Seamless Mobile Context: Towards High
Speed Event-based Middleware. In: Proceedings of the Fifth IEEE Consumer
Communications & Networking Conference (CCNC), Las Vegas, Nevada,
January (2008).
3. Hopfgartner, F., Jose, J. On User Modeling for Personalized News Video
Recommendation. In G.-J. Houben et al. (eds): UMAP 2009, LNCS 5535, 403-
408, 2009.
4. Loeb, S., Architecting Personalized Delivery of Multimedia Information.
Communications of the ACM, December 35(12) 39-48 (1992).
5. Loeb, S., Falchuk, B., and Panagos, E. The Fabric of Mobile Services: Software
Paradigms and Business Demands. Wiley (2009).
6. Panagos, E., Loeb, S., Falchuk, B. Actionable User Intentions for Real-Time
Mobile Assistant Applications. In: Workshop on Personalization in Mobile
Pervasive Computing, in User Modeling, Adaptation and Personalization
(UMAP), Trento Italy, June (2009).
7. Yu, Z., Yu, Z., Zhou, X. Nakamura, Y. Handling Conditional Preferences in
Recommender Systems. In: Proc. Of the IUI Conference, Sanibel Island,
Florida, 407-411 February (2009).
24
An Adaptive Game to Train Users with Special Needs
to Make Decisions: Using Public Transportation
EsteIania Martin
1
, Silvia Calvo
1
, Rosa M. Carro
2

Escuela Tcnica Superior de InIormatica, Universidad Rey Juan Carlos,
Mostoles, Madrid, Spain
esteIania.martinurjc.es, s.calvogalumnos.urjc.es

2
Escuela Politcnica Superior, Universidad Autonoma de Madrid, Spain
rosa.carrouam.es
Abstract. During our liIe, we are involved in a huge variety oI activities and
have to make many decisions. However, not all the users have the same
capabilities. Decision-making, as well as the tasks to be perIormed daily, can be
trivial Ior some users and quite hard Ior others. This paper presents an adaptive
game whose goal is to train users with cognitive disabilities to use public means
oI transport. The situations posed to the users at each time are adapted
according to their personal abilities and needs, especially considering their
cognitive disabilities and their progress during the game. The game supports
training their ability to react and make decisions when traveling Irom one place
to another using public transportation.
Keywords: Adaptive games, users with special needs, public transportation.
1 Motivation
During our liIe, we are involved in a huge variety oI activities that we constantly
repeat associated to diverse contexts. Furthermore, during each day, there are many
decisions to be taken, both in regular situations and in unexpected ones. However, not
all the users have the same capabilities, especially people with special needs.
A set oI considerations and recommendations to design activities Ior people with
cognitive disabilities are presented in |2| and |3|. People with cognitive limitations
have troubles to comprehend and perceive the environment. They have diIIiculties
with the transIer and consolidation oI learning; Ior this reason, new concepts to learn
and problems to solve should be contextualized in daily situations to help learners to
transIer knowledge to the real world. Other limitation relates to literacy skills. Users
have troubles to process inIormation while they are reading complex texts or listening
to too much inIormation. Furthermore, they need the concepts to be presented in a
repeated and Ilexible way in order to assimilate them. Finally, they have attention
diIIiculties and can Ieel Irustrated and disoriented iI the activities are monotonous or
their level oI diIIiculty is higher than their capabilities to solve them. Some oI the
main categories oI Iunctional cognitive disabilities include deIicits or diIIiculties with
memory, literacy skills, attention and problem solving.
25
Regarding public transportation systems, they are among the most complex
systems Iound in our modern society |6|. People with cognitive disabilities have
diIIiculties when they are using public transportation because a series oI high-level
activities that include planning, waiting and moving are involved.
2 Game Basis
Computers oIIer valuable assistance to people with special needs, including those
with physical and cognitive limitations. They oIIer new opportunities to learn, share
inIormation and gain independence |10|. Adaptation methods and techniques can
contribute to adapt existing soItware to better suit the user needs |12|. The activities
to be perIormed must be structured and adapted according to the abilities and needs oI
each user. The navigational guidance between activities should be direct, so they do
not lose attention Iocus. Problems presented must be contextualized in real situations
to help the transIer oI knowledge and the abstraction oI the concepts learnt. It is
important to repeat the same tasks several times in order to contribute to concept
assimilation. Instructions oI activities must be clear and easy to read. Regarding
contents, it is better to present inIormation in a visual way rather than using texts.
Combining diIIerent types oI multimedia material is useIul to increase user
motivation and attention and helps users to process inIormation.
The game presented in this paper can be used with diIIerent purposes. Players will
learn about the Madrid underground network while planning trips between diIIerent
stations. Furthermore, this game contributes to train users to make decisions when
unexpected situations happen. It will help users to achieve more independence in their
daily liIe. Users able to accomplish simple tasks regarding travel planning can also
learn to plan more eIIicient and eIIective trips.
Activities oI diIIerent complexity are proposed to distinct types oI users, taking
into account issues such as the time a trip is expected to last or the number oI stations
and the transIers involved. The game adapts the situations posed to the users at each
time according to their personal abilities and needs, especially considering their
cognitive disabilities and their progress during the game. Cognitive disabilities are
measured by using the questionnaire WAIS III |19| using the global estimation oI
individual cognitive Iunctioning, as well as the Iour index scores that represent the
major components oI intelligence. These Ieatures constitute the user model, which is
updated while the player is interacting with the game. The game adaptation is done by
means oI rule-based mechanism.
The game is structured in Iour levels. Each level has a set oI questions associated.
II a user answers a certain number oI questions correctly, the game allows him to
continue with the next level. The minimum percentage oI correct answers required to
allow users to go Irom a level to the upper one is 80. This number has been chosen
in order to users cannot become Irustrated or disoriented.
Questions can be classiIied according to their goal. Some oI them are intended to
help the user to become Iamiliar with the subway network; others are Iocused on
making the user to apply his knowledge oI the network. Questions describe realistic
situations, either usual or unusual, in which the user has to propose a solution.
26
In general, each question has two or more answers associated and only one oI them
is correct. Questions Ior the Iirst level include two potential answers. Questions Ior
upper levels oIIer more potential answers to increase the level oI diIIiculty when
taking the decision about which one is correct. When users are answering to a
question can consult the underground map, always available at the interIace. The
Ieedback given to the user is always positive. In the upper levels, one question may
have several potential answers. In these situations, the game takes into account
whether the user gave the optimal solution.
3 Conclusions and Future Work
The adaptive game presented should contribute to train users with cognitive
limitations to use public means oI transport. This game attempts to increase user
personal autonomy and to improve abilities such as user selI-control and decision-
making. With this game, players Iirstly can get Iamiliar with the underground Madrid
network. When interacting with it repeatedly, they are expected to become trained to
select the best route between two places. Furthermore, their ability to take decisions in
unexpected situations is expected to improve. With these purposes, questions are
adapted to the level in which the player is at each time and his abilities, in order to
avoid Irustration or disorientation. The teachers participating in the program
PROMENTOR |20|, who work daily with people with disabilities, have validated
both the game structure and the design oI the user interIace. We plan to evaluate the
game with all the users participating in PROMENTOR within the next months.

Acknowledgments. This research has been supported by the Spanish Ministry oI Science and
Education, project HADA (number TIN2007-64718).
References
2. Aubrey, C., Dahl, S.; A review oI the evidence on the use oI ICT in the Early Years
Foundation Stage. Early Childhood Research Unit, Institute oI Education, University oI
Warwick (2008)
3. Friedman, M.G., Nelson B.D.: Web Accessibility Design Recommendations Ior People with
Cognitive Disabilities. Technology & Disability, vol. 19(4), pp. 205--212. IOS Press (2007)
6. Carmien, S., Dawe, M., Fischer, G., Gorman, A., Kintsch, A., Sullivan, James F.: Socio-
technical environments supporting people with cognitive disabilities using public
transportation. ACM Transactions on Computer Human Interaction, vol. 12(2), pp. 233--
262. ACM. New York, USA (2005)
10.Special Technological Possibilities Ior Students with Special Needs. Learning & Leading
with Technology, vol. 29, number 3, pp. 32--39 (2001)
12.Brusilovsky, P.: Adaptive hypermedia. In: Kobsa, A. (eds.) User Modeling and User
Adapted Interaction, Ten Year Anniversary Issue, vol. 11 (1/2), pp. 87--110 (2001)
19.Wechsler, D.: The Wechsler Adult Intelligence Scale. San Antonio, TX: Psychological
Corporation (1997)
20.PROMENTOR program. Universidad Autonoma de Madrid and Fundacion PRODIS.
27
Towards Adapting Group Activities in Multitouch
Tabletops
Estefana Martn
1
and Pablo A. Haya
2

1
Universidad Rey Juan Carlos, c/Tulipn s/n, Mstoles, Madrid, 28933, Spain
estefania.martin@urjc.es
2
Universidad Autnoma de Madrid, Campus Cantoblanco, Madrid, 28049, Spain
pablo.haya@uam.es
Abstract. Face-to-face collaborative learning has been applied in traditional
classrooms since the 70s. Nowadays, many teachers use electronic devices such
as computers, digital backboards, tablet PCs, etc. as everyday tools in their
classrooms from kindergarten to University. In recent years, the use of
multitouch surfaces has increased in our society. This paper presents a proposal
to adapt group activities that will be performed in multitouch tabletops.
Keywords: Adaptation, Group Activities, Multitouch Tabletops.
1 Motivation
The origins of group activities are based on the real world. In our life, we are
continuously interacting inside groups, e.g. in the family life or with our friends. Our
personal identity stems from the way of perceiving and interacting with other group
members. Within the group we learn how to behave, to think, and to educate
ourselves from our interaction with the rest of the members [1].
Face-to-face collaborative learning has been applied in traditional classrooms since
the 70s [2]. Different studies arose trying to apply collaborative learning technologies
to pupils of diverse ages. These experiences pointed out that the learning process is
not only the knowledge acquisition, but also includes the detection of missed
information and inconsistencies [3]. Students share knowledge and acquire new
knowledge in collaborative learning [4]. Furthermore, collaboration has great benefits
such as to promote the interaction and the familiarity among students and teachers, to
facilitate the development of reasoning skills, and to increase the students motivation
and participation [5].
The inclusion of collaborative activities in the classrooms have a great variety of
possible configurations depending on many variables, such as size of the collaborative
group, the different criteria for grouping students, the nature of the collaborative
activity, how long the students must stay at the same group, etc. These aspects could
be considered to adapt the group activities to the students.
Nowadays, many teachers use computers, digital backboards, tablet PCs, etc. as
everyday tools in their classrooms from kindergarten to University. Likewise, in the
28
recent years, the use of multitouch surfaces has increased in our society. We believe
that multitouch surfaces, specifically multitouch tabletops, will be inside the
classrooms in the next few years. Based on this premise, next section presents an
approach to adapt educational group games for multitouch tables detailing the
adaptation features considered in the user model. Finally, conclusions and future work
are presented.
2 Proposal
When users are interacting with multitouch tables, they stay around them. They
interact through natural gestures using their fingers. For example, they can press over
the object selected to click the object or drag and drop it. We use FLING (Flash
Library for Interpreting Natural Gestures) [6] to capture the users gestures while
they are interacting with the tabletop. The users' interactions should be different
depending on the nature of the activity or its goal. Other aspects that could influence
in the development of the activities are the number of the users, their physical location
around the table, their level of knowledge, their actions, the number and the size of
the objects over the table, etc. In this way, all these aspects build the user model and
they will be taken in mind for adaptation purposes. The user model should be updated
according to the users interactions with the tabletop.
The next step is to create the group activities that will perform using multitouch
tabletops. The teacher is the responsible to create them, to decide their type
(competitive or collaborative), to provide their contents and to specify the adaptation
features. In order to study the adaptation of group activities to users when they are
interacting with the multitouch table, we propose a card-based game. We choose card
games since they have three broad benefits: a) they are easy to understand; b) cards
can be directly manipulated; and c) many different games can be generated changing
the contents but preserving the rules. In our case, the number of participants varies
from 2 to 4 students around the table. The goal is to associate cards with similar
meaning. For instance, a digit card with a photo card that displays the number of
objects represented by the digit. At the beginning, a variable number of cards are
distributed in random positions and orientations on the tabletop. These cards are
called free cards. All participants have a configurable number of target cards in front
of themselves. They have to associate their target cards with the free cards on the
tabletop. Each target card has to have at least one free card that matches it, and, it is
not possible that a free card does not match any of the target cards. When a free card
is correctly associated to a target card, it disappeared from the tabletop. The game
ends when there are no free cards left on the tabletop. Positive and negative points are
assigned according to the number of correct and wrong associations.
The game has two modes: cooperative and competitive. In the collaborative mode,
participants share a common goal and they obtained a unique score. An association
can be completed by any of the participants even if the target card does not belong to
her. Cards are selected by finger pressing on each card of the pair. On the contrary, in
the competitive mode, each participant has her own goal. Points are obtained
29
individually. In this case, the association is fulfilled when the free card is positioned
on top the correct target card, i.e., the free card has to be drag and dropped.
In both modes, participant's target cards can be different or repeated. This affects to
difficulty level of the game, as when target card are repeated, the game is easier in the
collaborative mode, but harder in the competitive one. Additionally, the rest of the
previous adaptation features considered at the beginning of this section should be
included if the teacher desires. Specifically, the difficulty level can be considered by
means of the card content. For instance, in a digit card example, the digit target card
distribution affects to the activity complexity. On one hand, target cards having lower
numbers are easier than higher one (e.g. 1 and 2 vs. 8 and 9). On the other hand,
having similar numbers are more difficult than different one (e.g. 5 and 6 vs. 1 and 7).
3 Conclusions and Future Work
Multitouch tabletops provide an environment where students can interact performing
different group activities in an enjoyable way. Non-adaptive applications to
multitouch tabletops have already been developed and tested [7]. However, the
students interactions, their personal features, the goal and the nature of the own
activities could be influence in the development of them. So, it is necessary to bear in
mind these aspects to adapt the activities to the users. The proposal presented in this
paper defines some adaptation features that will be considered for collaborative
learning using a multitouch tabletop. Different adaptive games would be generated
depending on the adaptation features specified by the teacher. Similar games should
be designed with letter cards or with symbol cards too. Furthermore, we want to study
what adaptation features will be more relevant when users are performing this type of
game with these tabletops.

Acknowledgments. This work has been partially funded by the following projects:
HADA (TIN2007-64718), VESTA (TSI-020100-2009-828) and eMadrid
(S2009/TIC-1650).
References
1. Johnson, D.W., Johnson, F.P.: Joining together: group theory and group skills. Pearson
Allyn & Bacon (2002)
2. Slavin, R.E.: Cooperative learning. Review of Educational Research 50, 315--342 (1980)
3. Vygotsky, L.S.: Mind in society: The development of higher psychological processes.
Cambridge MA: Harvard University Press (1978)
4. Piaget, J.: Piaget's theory. In P. Mussen (ed.). Handbook of Child Psychology. 4th edition.
Vol. 1. New York: Wiley (1983)
5. Zurita, G., Baloian, N., Baytelman, F., Farias, A.: Developing Motivating Collaborative
Learning Through Participatory Simulations. Computational Science ICCS 2007. LNCS,
vol. 4488, pp. 799--807. Springer, Heidelberg (2007)
6. Flash Library for Interpreting Natural Gestures (FLING): http://amilab.ii.uam.es/fling
7. ShareIT Project: http://www.shareitproject.org/
30
Enriching TV Experience: TV-viewing Support System
Considering Both Individual and Family Preferences
Chihiro Ono
1
, Kazushi Ikeda
1
, Gen Hattori
1
, Kazunori Matsumoto
1
and Yasuhiro
Takishima
1
,

1
KDDI R&D Laboratories, Inc. 2-1-15 Ohara Fujimino, Saitama, 356-8502, JAPAN,
{ono, kz-ikeda, gen, matsu, takisima}@kddilabs.jp
Abstract. We propose a TV viewing support system which realizes a novel TV-
viewing style whereby family members interactively enjoy collectively preferred
TV programs and TV-related information comfortably at the same time by extensive
cooperation between a set-top-box (STB) and cellular phones of each member. Here,
1) a STB connected to the TV automatically detects family members by
communicating with cellular phones via a Bluetooth interface, 2) automatically
searches web pages synchronized with the progress of the TV program by analyzing
TV closed captions by scenes, 3) recommends TV programs and TV-related
information such as web pages and advertisements by considering both individual
preferences for cellular phone screens and family preferences for TV screen using
both individual and family usage histories. We confirm the effectiveness of our
proposed recommendations thorough experiments on TV program recommendation.
Keywords: recommender systems, group recommendation, statistical modeling.
1 TV-viewing Support System
Recently, the habit of users retrieving information on interesting shops and
restaurants, etc. via cellular phone and PC while viewing TV programs has increased,
and reports indicate 68% of Internet users do so in Japan, as opposed to 33% in the
United States. To make the user's TV viewing experience much more comfortable, we
propose a TV viewing support system which realizes a novel TV-viewing style
whereby family members enjoy collectively preferred TV programs and TV-related
information comfortably and at the same time. Fig. 1 shows an overview of the TV-
viewing Support System. Here, we assume a scene where family members are
viewing TV via their own cellular phones in the living room. Each member uses a
cellular phone as a side-screen for viewing preferable TV-related information (Fig. 1.
bottom right) as well as an IrDA TV remote controller (Fig. 1 bottom left). On the TV
screen, by automatically balancing the preferences of family members who are
currently viewing TV program, a list of TV-related web pages (right of the TV screen
in Fig. 1) and TV-related advertisements/movies (bottom left of TV screen in Fig. 1),
which may fit the balanced family preference, are displayed.
31

Fig. 1. Screenshots of TV-viewing Support System
The following are issues and solutions to realize the TV-viewing Support System:
(1)Detection of users viewing TV and history storage To detect users watching TV,
we use a Bluetooth interface (serial Port Profile) so that a STB could search nearby
cellular phones, because the number of cellular phones with Bluetooth interfaces is
increasing in Japan. To understand family preferences, TV viewing histories and web
pages played on the TV screen are stored by the unit of the members viewing TV,
rather than as individual units. Conversely, histories on cellular phones are treated as
individual histories and forwarded to a recommendation server via the mobile internet.
(2)Automatic collection and feature extraction of candidate TV-related
information To search web pages synchronized with the progress of the TV program,
we use a closed caption of TV program and analyze it to obtain important keywords
that express the features of the scene by scenes. Candidate information is collected via
Web search engines using extracted important keywords, following which the features
of web pages are extracted to be evaluated for selection for recommendation.
(3)Recommendation method considering both personal and family preferences
For modeling preferences of individual users, by collecting user attributes and
content-selection histories from many users, we build both positive and negative
preference models in the form of a Bayesian network model. Here, the preference
model is formalized as modeling the conditional probability distribution P (C |U, S) to
predict the content value C from user U and situation S. At constant intervals, in order
to adapt to each individual user and each combination of families, newly collected
personal histories are combined by weight with the overall history data and the
conditional probability table of the Bayesian network model is renewed.
A family with more than two members has more than two groups that are
combinations of family members, each of which has a group-type which reveals the
characteristics of the groups. As a group-type, we have following 4 types. Here, F(x,
) is a predicted evaluation of the content x with a threshold ,
i
is a weighting
parameter of predicted evaluation of each user i, x
i
is a predicted evaluation of user i,
and
i
is a threshold of user i used in Mixed LOA with priority model.
(1) Simple additional model: Recommend items by making a list of predicted
evaluations with the summation of the predicted individual evaluations:
32
(2) Priority model: Recommend items by making a list of predicted evaluations by
summation of the predicted individual evaluation by further prioritizing major
decision-making individuals. Here logistic regression is used based on preliminary
experiments.


(3) Mixed LOA (Level of Aspiration) with priority model: Recommend items
based on the assumption that the more acceptable users, the better. Here, we define
acceptable as a predicted evaluation exceeding a certain threshold and determine the
threshold for each combination of family members using history data of the target
combination.
(4) Different character model: Positive and Negative preference models are
established using the attributes of groups such as the number of members, with child
or not, etc., and situation attributes, and the positive and negative histories of the
target combination.
When deciding a group-type for each combination of family members, we calculate
the correct rate using training data sets and decide the type to the highest one.
2 Evaluation of preference models
The proposed models are evaluated for the TV program recommendation task using
history storage. The correct answer rate was calculated from among seven programs
telecast to an isochronal belt considering a correct answer to be when a program that
seemed interesting was predicted, and the program actually viewed came in more than
2nd place. Here, the correct answer rate for random selection was 28.7% (=2/7*100).

Table 1. Average correct rates
before adaptation after adaptation
Individual model 49.0% 52.9%
Considering the case where all the group members
are allocated to the simple additional model
45.6% 53.2%
Group models

Average of four models 55.3% 62.4%
Simple additional model 52.1% 57.2%
Priority model 59.0% 66.9%
Mixed LOA with priority model 60.9% 77.3%
Different character model 40.8% 49.5%
As shown in Tab. 1., the proposed method which considers the group-type (55.3%)
was far better than the simpler case where the group-type was not considered but
regarding all the groups allocated to the simple additional model (45.6%). As for the
effectiveness of the adaptation, the correct rate of the preference estimation improved
from 55.3 to 62.4%by adaptation using only 3 week histories of target users.
^ ` ^ ` ^ `
C B A
x x x x F u u u 1 1 1 ) , ( G
)) ) ( ) (( exp( 1
1
) , (
C x x
x F
B B A A
u u

E E
G
^ ` ^ `
^ `

) 0 , ) ((
) 0 , ) (( ) 0 , ) (( ) , (
! u
! u ! u
not x x if
not x x if not x x if x F
C ABC C C
B ABC B B A ABC A A
G E
G E G E G
33
Patient Condition Modeling in Remote Patient
Management: Hospitalization Prediction
Mykola Pechenizkiy
1
, Aleksandra Tesanovic
2
, Goran Manev
1,2
,
Ekaterina Vasilyeva
1
, Evgeny Knutov
1
, Sicco Verwer
1
, and Paul De Bra
1
1
Department of Computer Science, Eindhoven University of Technology
P.O. Box 513, NL-5600 MB, Eindhoven, the Netherlands
{m.pechenizkiy,e.vasilyeva,e.knutov,s.verwer}@tue.nl, debra@win.tue.nl
2
Philips Research Laboratories
High Tech Campus 37, 5656 AE Eindhoven, the Netherlands
{goran.manev,aleksandra.tesanovic}@philips.com
Abstract. In order to maintain and improve the quality of care with-
out exploding costs, healthcare systems are undergoing a paradigm shift
from patient care in the hospital to patient care at home. Remote patient
management (RPM) systems oer a great potential in reducing hospital-
ization costs and worsening of symptoms for patients with chronic dis-
eases, e.g., heart failure and diabetes. Dierent types of data collected
by RPM systems provide an opportunity for personalizing information
services, and alerting medical personnel about the changing conditions
of the patient. In this work we focus on a particular problem of patient
modeling that is the hospitalization prediction. We consider the prob-
lem denition, our approach to this problem, highlight the results of the
experimental study and reect on their use in decision making.
1 Introduction
Chronic diseases are the leading cause of death and healthcare costs in the de-
veloped countries. Healthcare systems are undergoing a paradigm shift from
patient care in the hospital to the patient care at home [3]. It is believed that
RPM systems, by providing adequate patient monitoring, instruction, educa-
tion and motivation (all of which can be done outside of the hospital) facilitate
normalization of the patients conditions and prevent re-hospitalization.
Recently, a possible architecture of the next generation of personalized RPM
systems was introduced, and a general process of knowledge discovery from RPM
data, leading to identication of potentially useful features and patterns for
patient modeling and construction of adaptation rules, was considered [2].
In this paper we focus on the problem of timely patient hospitalization pre-
diction, particularly Heart Failure Hospitalization (HFH). Currently, domain
experts are using manually designed triggers that should trigger an alarm in the
case of possible HFH. Our study shows that with the intelligent data analysis
approach for patient modeling, which utilizes information spread across dier-
ent data sources, it is possible to learn predictive models that are more accurate
than the expert-authored triggering rules (with statistical signicance).
34
2 Hospitalization prediction
The problem of HFH prediction can be dened in the following way: based on the
available data about a patient at moment t
i
cast a prediction (and raise an alarm
if deemed necessary) on a daily basis whether the hospitalization for this patient
is likely to occur within the next 14 day period, t
i
. . . t
i+14
. Figure 1 illustrates the
timeline of data availability used by a domain expert or an automated classier
for facilitating this decision making.
Fig. 1. Hospitalization prediction for the following 14 day window.
At the time of enrolment (t
0
) of a patient, complete medical history data
(corresponds to H features) is recorded. A record may contain dozens of elds
providing dierent information such as information related to previous hospital
admissions, existence of valve diseases, evidence of coronary diseases, arrhyth-
mias, devices implanted, etc. During a monthly phone contact (MC
j
) patients
are asked to assess quality of life (QoL) symptoms (S features), and report addi-
tional data such as disease and non-disease medication (or medication change),
number of visits/contacts (at home, by phone, at the oce, at the clinic) in the
last month. The patients are monitored on a daily basis regarding their vital
signs such as weight or blood pressure (source for constructing D features).
Figure 2 shows our approach of constructing positive training instances, i.e.
the case when HFH took place. We nd a day on which HFH has occurred (t
h
),
then take the 14 days window [t
h14
, t
h
) to compute features related to daily
measurements. It should be noticed that data for computing these features may
include days outside this two week window.
Fig. 2. Forming of a positive (hospitalization tool place) training instance.
35
3 Experimental study
We performed a quantitative evaluation of our approach on an extract from the
TEN-HMS dataset [1] containing information about 426 patients with cardio-
vascular diseases, 43 of which had at least one HFH.
Our experiment setup consisted of two major steps. In the st place we ap-
plied dierent classication techniques, including e.g. support vector machines
(SVM), decision trees (J48), and rule-based learners (JRip)
3
. By means of cross-
validation on the training data, we searched for and xed the best parameters for
each classication technique and best feature set from S, H and D groups of fea-
tures. Then, the selected classiers were compared against individual triggering
rules on the testing data. All the learnt classiers were statistically signicantly
more accurate (about 10% on average) than any of the individual triggering
rules according to paired t-test with respect to Youden index (YI) that regards
true positive rate (TPR) and false positive rate (FPR) as equally important. J48
showed the lowest FPR, yet having slightly lower YI than SVM and JRip.
4 Conclusions and further work
In this paper
4
we presented a general approach for modeling patient state from
historical data of dierent kinds, including vital signs, system usage, medical
history and regular interviews and questionaires. We illustrated the potential of
our approach on the example of the HFH prediction problem by providing the
results of an experimental study with the data from a real clinical trial.
Our work laid the foundation for facilitating better personalization and alert-
ing services in RPM systems, and we plan to continue working in this direction,
particularly improving HFH prediction. We plan to make use of the educational
data, motivational messages and other feedback provided to the patient by an
RPM system or medical personnel, to obtain reliable and up-to-date information
about the symptoms, and to make our prediction approach context-aware.
Acknowledgments. This research is partly supported by EU HeartCycle and
KWR MIP (Medical Information Processing) projects.
References
1. J. Cleland, A. A. Louis, A. Rigby, U. Janssens, and A. Balk. Noninvasive home
telemonitoring for patients with heart failure at high risk of recurrent admission
and death. Journal of American College of Cardiology, 45(10):16541664, 2005.
2. A. Tesanovic, G. Manev, M. Pechenizkiy, and E. Vasilyeva. ehealth personaliza-
tion in the next generation rpm systems. In Proceedings 22nd IEEE International
Symposium on Computer-Based Medical Systems, pages 18. IEEE Press, 2009.
3. H. Wang. Disease management industry and high-tech adoption. An Industry report
from parks assiociates, Parks Associates, 2008.
3
We used WEKA 3.6 data mining toolkit, www.cs.waikato.ac.nz/ml/weka/
4
The extended version is accessible at www.win.tue.nl/
~
mpechen/projects/rpm/
36
Clustering of Web Users Using the Tensor Decomposed
Models
Rakesh Rawat, Richi Nayak,YueIeng Li
Faculty oI Science and Technology,
Queensland University oI Technology, QLD 4001, Brisbane Australia.
r.rawat,r.nayak,y2liqut.edu.au}
Abstract. We propose to use the Tensor Space Modeling (TSM) to represent
and analyze the user`s web log data that consists oI multiple interests and spans
across multiple dimensions. Further we propose to use the decomposition
Iactors oI the Tensors Ior clustering the users based on similarity oI search
behaviour. Preliminary results show that the proposed method outperIorms the
traditional Vector Space Model (VSM) based clustering.
Keywords: Tensor Space Modeling, Clustering, Web data mining.
1. Introduction
Web server log data is a rich source oI inIormation about the users browsing habits.
This data can be used Ior clustering oI similar search behavior, and subsequently used
Ior eIIective personalization |1|. Considering the multi-dimensional nature oI the
users search data, we propose to use Tensors to model the Web server log data, and
subsequently utilize the best rank approximations inIerred using the tensor
decomposition techniques such as PARAFAC (Parallel Iactors) |2| and Tucker |3| Ior
clustering the Web users. Tensors have been used widely in chememometrics due to
their multi dimensional data modeling abilities and inIerence capabilities |4|,|5|,|9|.
In spite oI popularity oI tensors in multi dimensional data modeling techniques,
seldom tensor based clustering is discussed, except a couple oI research |6|,|9|.

2. Method

Searches made by each user on a Web site are processed Ior identiIying search
sessions. Each search session can be represented by a set oI interest vectors Ior each
user. Let an interest vector be denoted as
1 2
( , ,...., )
N P
T F F F = that contains P search
query components or dimensions (like make, model, cost etc.). Let
M N
LX T denote an
interest vector oI user
M
X showing a particular search T
N
made by the user. Let
1 2
| , .. |
M M M M O
,8 LX T LX T LX T = be a matrix containing all the interest vectors oI
M
X
.
For a
user
M
X , two interest vectors
M N
LX T

and
M O
LX T

are considered same when
, iI ( ) ( ) ( )}.
M N M O P N P O
LX T LX T F F N O = . =

Using the ( ) WI LGI - weighting scheme, a

37
weight value is calculated Ior each interest vector oI each user as
1 log( ) ,
1'
7)
')
| |
+
|
\ .
where 7)Frequency oI interest vector
M N
LX T

Ior user
M
X ,
1'Total number oI interest vectors oI the user
M
X and ') Number oI interest
vector
Q N
LX T

oI all users in the dataset.

A sparse tensor 7 is modeled with the
normalized weight ( )
S
: values Ior each interest vector oI all users denoted as
1...
( , ) }, 7
S Q S
T X : =

where S

is the total number oI unique interest vectors searched
by all users and Q is total number oI users in the search log Iile. Next the tensor
model is decomposed in order to capture the multi linear structure in data by
analyzing the Iactors in each mode. The component matrices
1 2 1
, ..
P+
A A A are
obtained aIter PARAFAC |2| and Tucker |3| decomposition and represent
decomposed values oI each mode
1 2 1
..
P
0 0 0
+
respectively. Each component
matrix is oI dimension
L U
A where L is the number oI ways in a mode
P
0 , and U
is the value oI best rank approximation oI the tensor. We have taken the
1 P+
A matrix
as clustering input, as it represents the users dimension and subsequently it depicts
the correlations with other users based on the multiple Iactors oI the tensor.

3. Evaluation

This research uses the server log data collected Irom a popular car sales Website
1

in Australia which allows users to do parameterized searches. The object oI search is
a car, and the parameters oI search query are Car Make, Model, Car Body type, Car
Cost and Car Search type. The objective was to cluster users based on similar search
behavior. A tensor with modes 7
0DNH 0RGHO %RG\W\SH &RVW 6HDUFK7\SH 8VHUV
Q

e\ is created.
Sample dataset had 880 search records oI the 25 users. The numbers oI ways in each
dimension were
13 38 16 14 4 25
7

e\ . Decomposition was achieved using PARAFAC
and Tucker methods |7|. The decomposed Iactor values oI model represented by
component matrix
1 P+
A with best rank approximation value in the range 1 to 6 were
then clustered using the K-means and EM (Expectation Minimization) |8|. Clustering
was considered good when two users search behavior was similar, meaning that they
had searched Ior similar search parameters and were clustered together. Fig. 1 shows
that average and overall perIormance oI clustering using TSM models is much better
than clustering using the Vector Space Model (VSM).

Fig. 1. Average Clustering dataset 1-F-Score Ior 25 Users.

1
Due to privacy issues, we will not reveal the exact address and detail oI the Web site.
38
4. Discussion

It can be seen that when data is multidimensional and consists oI inter related
components, two way clustering using VSM may not be as accurate as an Q way
clustering approach. In a multi dimensional dataset such as Web search data, multi
way analysis is able to extract hidden structures or patterns that capture underlying
correlations between various dimensions. This underlying inIormation oI the data
may not be captured accurately or identiIied uniquely by two-way analysis methods.
Such methods e.g. the Iactor models, suIIer Irom rotational Ireedom unless speciIic
constraints such as statistical independence, orthogonality, etc. are enIorced |5|. Since
matrices have only two dimensions, there arises a need to represent such data multi
linearly so that the underlying relationship between each dimension independently or
in correlation to others can be analysed in a more eIIicient manner. The other
noteworthy observation is that Ior tensor decomposition methods like PARAFAC or
Tucker, as the rank oI tensor best rank approximation increases the clustering
perIormance decreases. It was Iound out that optimal clustering results were obtained
when the rank is nearly halI the number oI dimensions (Fig. 1). As the rank increases
beyond this the perIormance decreases and decomposition at certain ranks may equal
or be lower to the clustering results as obtained by standard vector space models. A
reason may be the loss oI inter related inIormation between Iactors Ior higher ranks.

ACKNOWLEDGEMENT:

We would like to thank the Smart Services Cooperative Research Centre Ior partly
Iunding this project.

References


|1| Mobasher, B., "Data Mining Ior Web Personalization," in The Adaptive Web. vol. 4321, A. K.
P. Brusilovsky, and W. Nejdl (Eds.), Ed., 2007, pp. 90-135.
|2| Harshman, R. A., "Foundations oI the PARAFAC procedure: Models and conditions Ior an"
explanatory" multi-modal Iactor analysis," UCLA working papers in phonetics, vol. 16, p. 1,
1970.
|3| Tucker, L. R., "Some mathematical notes on three-mode Iactor analysis," Psychometrika, vol.
31, pp. 279-311, 1966.
|4| Kolda, T. G. and Bader, B. W., "Tensor decompositions and applications," Technical Report
SAND2007-6702, Sandia National Laboratories, Albuquerque, NM and Livermore, CA, ,
November 2007.
|5| Acar, E. and Yener, B., "Unsupervised multiway data analysis: A literature survey," Technical
report, Computer Science Department, Rensselaer Polytechnic Institute, Troy, NY. , 2007.
|6| Jegelka, S., Sra, S., and Banerjee, A., "Approximation Algorithms Ior Tensor Clustering," in
LNCS. vol. 5809/2009: Springer, 2009, p. 368.
|7| Kolda, B. W. B. a. T. G., " MATLAB Tensor Toolbox Version 2.2,," in
http://csmr.ca.sandia.gov/~tgkolda/TensorToolbox/, , January 2007.
|8| Ian H. Witten, E. F., Data Mining: Practical machine learning tools and techniques, 2nd Edition
San Francisco: Morgan KauImann, 2005.
|9| Kroonenberg, P M., Applied Multiway Data Analysis, Wiley-Interscience, 2008.

39
Towards Incorporating Confidence Measure
Into User Modeling
Michael Yudelson
1
, Peter Brusilovsky
1
, Antonija Mitrovic
2
, Moffat Mathews
2


1
University of Pittsburgh, School of Information Sciences,
135 North Bellefield Ave. Pittsburgh, PA 15260, USA
2
University of Canterbury, Department of Computer Science and Software Engineering
Private Bag 4800, Christchurch 8140, New Zealand
{mvy3, peterb}@pitt.edu, {tanja.mitrovic, moffat.mathews}@canterbury.ac.nz
Abstract. User model is a corner stone of every user-adaptive system. Since the
emergence of the field of user modeling a great number of approaches to
modeling users were proposed. All of these approaches, in one way or the other,
capture various user traits, aspects of the target domain and task, etc. Several
researchers working on user modeling in interactive learning systems found
that, despite good predictive power of the models, they are often very optimistic
about the user characteristics. Nearly all user models lack a form of confidence
measure that would provide a conservative assurance that the models view of
the user is accurate and the values are not inflated.
In this work we are discussing a candidate for such confidence measure that is
based on a Wald-Wolfowitz test for randomness (cf. [7]). We evaluate the
usefulness of the suggested confidence measure by incorporating it into two
different approaches to building user models and testing on a real-world student
activity data. Our estimates show that utilizing confidence helps reliably
improve the accuracy of both user modeling methods in question with respect to
a set of widely accepted user model accuracy measures.
Keywords: User Modeling, Confidence, Bayesian Knowledge Tracing,
Educational Data Mining.
1 Modeling Students Knowledge And Confidence
Many overlay student modeling methods keep a parameters for each of the modeled
skills. For example, in Bayesian Knowledge Tracing [3] four parameters are used to
estimate students mastery of each skill: probability of knowing the skill, probability
of learning the skill next time it is practiced, probability of making a careless mistake
(slip), and probability of correctly guessing (guess). In a parameterized version of
CUMULATEs asymptotic modeling algorithm [2] three are used: knowledge level of
the skill, speed of knowledge level growth, and decay factor for [over-]practicing the
skill when attempting the same problem where this skill is involved. Estimations of
such parameters are computed based on the students history of practicing the skill
when interacting with an educational system.
40
In statistics, the estimations are often accompanied by confidence intervals. The
size of the confidence interval shows close is the estimate to the true parameter value
(the smaller the interval, the closer we are) [4]. In association rule mining, each found
relation has two values: support the amount of data that advocate the existence
relation, and confidence is the certainty that the relation found in fact exists [6].
While these two confidence measures have different application, assumptions, and are
computed differently, they both offer a second opinion, a reliability measure. In
overlay student modeling, until recently, there were few to none approaches that
would capture both the estimates and the confidence value of the assessed parameters.
Recently, researchers realized that despite promising student model accuracy levels
obtained on internal validity tests using system logs, the external validity tests
accuracy (obtained using for example paper-based pre- and post-tests) is visibly
lower. Student model is often found to be overly optimistic about users skill mastery
levels [1]. Authors of [1] suggest a modification to a classical Bayesian Knowledge
Tracing (BKT) method in attempt to offset these deficiencies. They suggest using
additional contextual guess and slip parameters to compensate for unnecessary
inflation of the skill mastery estimate. While the proposed BKT modification has been
shown to tangibly improve models performance on external validation tests it is
specific to BKT approach to user modeling only. In addition, the added contextual
parameters do not directly correspond one to each skill.
In this work we are proposing a way to extend an arbitrary user modeling approach
with a confidence measure for each of the modeled skills. We are basing our measure
of confidence on Wald-Wolfowitz test for randomness [7]. Integrating this confidence
measure with two user modeling approaches shows significant improvements across
several model accuracy metrics.
2 Confidence Measure And Its Integration Into User Model
Wald-Wolfowitz test also known as runs test for randomness is used to test the
hypothesis that a series of numbers (a run) are random. In the context of user
modeling, we would like to know if a sequence of correct or incorrect applications of
a skill is random or not. Runs test here servers the purpose of a very conservative user
model that would keep our main user modeling approach in check.
The two approaches that runs test-based confidence was incorporated with are
Bayesian Knowledge Tracing (BKT) [3] and parameterized version of CUMULATE's
asymptotic algorithm (PCAA) [2]. Student logs of the intelligent tutoring system
SQL-Tutor (cf. [5]) from 2006-2007 were used to assess models performance. Logs
contained a total of about 10500 problem solution attempts of about 120 users from 3
classes that used SQL-Tutor. Both BKT and PCAA had individual parameters for all
282 skills contained in the datasets. These parameters were optimized using curve-
fitting procedure minimizing the mean squared error (MSE) of the modeling. MSE
was computed using actual and expected correctness of student problem solutions.
Confidence measure was computed the following way. If for the students
sequence of correct/incorrect skill application attempts runs test reject the null
hypothesis and the number of correct attempts is larger than the number of incorrect,
41
then the user is said to possess a reliable knowledge of that skill, otherwise the
knowledge of the skill is said to be unreliable.
Confidence and user model are integrated in a linear fashion. User model
prediction of student correct solutions and confidence measure were the explanatory
variables, while the actual problem-solving outcome was the binary response variable
of the binomial regression. Later the resulting predicted values were compared to the
user model predictions of correctness alone. Results show that across both user
modeling approaches, all three datasets, and several recognized user model accuracy
metrics (mean actual-expected error, area under ROC curve, correlation of actual and
expected accuracies, etc.), confidence-full models have a significant edge over
confidence-less ones (refer to poster for details).
3 Future Work
Work on a general confidence measure for student modeling is only in its beginning
stage. We only evaluated the impact of proposed confidence metric internal validity
tests. External validation using student post-tests is necessary to bring credibility to
the proposed approach. Wald-Wolfowitz test is not the only test that could be used as
a back-end of the confidence metric. Wilcoxon rank test for randomness of sign as
well as Kolmogorov-Smirnov test can potentially be used.
Behavior of confidence measure in general needs to be studied in depth. For
example, occurrence of student errors before and after confidence reached positive
value. Additional to the two discussed in this work, other user modeling methods
could be investigated for the usefulness of confidence measure incorporation.
References
1. Baker, R. S. J., Corbett, A. T., and Aleven, V. (2008). More Accurate Student Modeling
through Contextual Estimation of Slip and Guess Probabilities in Bayesian Knowledge
Tracing. In B. P. Woolf, E. Ameur, R. Nkambou, and S. P. Lajoie (Eds.), 9th International
Conference On Intelligent Tutoring Systems (ITS 2008), (pp. 406-415).
2. Brusilovsky, P., Sosnovsky, S. A., and Shcherbinina, O. (2005). User Modeling in a
Distributed E-Learning Architecture. In L. Ardissono, P. Brna, and A. Mitrovic (Eds.),
10th International Conference on User Modeling (UM 2005), (pp. 387-391).
3. Corbett, A. T. and Anderson, J. R. (1995). Knowledge tracing: Modeling the acquisition of
procedural knowledge. User Modeling and User-Adapted Interaction, 4(4), 253-278.
4. Fisher, R. A. (1956). Statistical methods and scientific inference. Oxford, England: Hafner
Publishing Co.
5. Mitrovic, A. (2003). An Intelligent SQL Tutor on the Web. Iinternational Journal of
Artificial Intelligence in Education, 13(2-4), 173-197.
6. Piatetsky-Shapiro, G. (1991). Discovery, Analysis, and Presentation of Strong Rules. In G.
Piatetsky-Shapiro and W. J. Frawley (Eds.), Knowledge Discovery in Databases (229-248).
AAAI/MIT Press.
7. Wald, A. and Wolfowitz, J. (1940). On a Test Whether Two Samples are from the Same
Population. The Annals of Mathematical Statistics, 11(2), 147-162.
42
Scrutinizing User Interest Models with
IntrospectiveViews
Fedor Bakalov
1
, Birgitta K onig-Ries
1
, Andreas Nauerz
2
, and Martin Welsch
1,2
1
Friedrich Schiller University of Jena
{fedor.bakalov | birgitta.koenig-ries}@uni-jena.de
2
IBM Deutschland Research and Development GmbH
{andreas.nauerz | martin.welsch}@de.ibm.com
Abstract. In this paper, we demonstrate the application of the Intro-
spectiveViews interface for scrutinizing user interest models. We illus-
trate how with this interface the user can get an overview of what the
system knows about her interests, obtain details about any interest,
and alter her user model.
1 Introduction
IntrospectiveViews [2] is an interactive visualization interface to user models.
Through this interface users can see what the system knows about them and
edit that information. These user models can then be used as the basis for
adaptation, e.g., for sorting news stories in a news portal. IntrospectiveViews
follows Shneidermans Visual Information Seeking Mantra [3]: overview rst,
zoom and lter, then details-on-demand. It oers users an overview over all
terms present in their interest models, it allows for zooming into dierent parts
of the model, ltering terms according to dierent criteria, and it will provide
details on demand.
This demo will showcase the usage of IntrospectiveViews as the user inter-
est model underlying an adaptive news portal. It will show how the user can
interact with IntrospectiveViews and how changes made to the user model via
IntrospectiveViews inuence the adaptation provided by the news portal.
2 Brief Introduction to IntrospectiveViews
IntrospectiveViews, shown in Fig. 1, is implemented as a Java Applet. It displays
user interests on a circular surface consisting of three colored rings. Positioning
of interests (terms) on the surface, namely the distance from the circle center, is
determined by the terms exact degree of interest. Here it means that the closer
a term appears to the center, the higher interest it represents. The interest
degree is also encoded in the terms font size. Each ring represents a certain
interest group. The color scheme of rings is chosen according to the hot-and-
cold metaphor, where hot, represented by red color, denotes interest and cold,
represented by blue color, denotes no interest. The colors between red and blue
43
Fig. 1. IntrospectiveViews Interface
Color screenshots and a screencast are available at http://www.minerva-
portals.de/research/introspective-views/
denote partial interest. The border areas of the rings are painted in a gradient
color to denote the fuzziness between the groups, i.e., uncertainty of interest
degree. User interests are grouped into circular sectors by type, i.e., the semantic
class they belong to (e.g. person, company, country, etc.).
The entire collection of interests can be zoomed in and out. The user can
navigate through the collection (e.g. in zoomed in view) by dragging the surface
in a respective direction. The interface supports a number of ltering options.
Terms can be ltered by type and by interest group. For instance, the user can
display only companies, people, and countries that she is highly interested in.
Also the user can obtain additional information about the terms. Left click on
a term will display the semantic relations to the related terms, whereas right
click will display the terms context menu through which the user can access
the terms description, get justication of her interest (either she has read many
documents about it, or her interest was propagated from other terms, or she
specied her interest herself), obtain the list of related documents, and view and
edit the privacy settings for the selected term(s).
In addition to viewing, the interface enables the user to edit her interests.
In order to change interest degree in a certain term, the user can simply drag
the term into the appropriate interest group (here represented by one of the
44
rings). In order to add new interest the user can double click in any place on the
surface and select the term from the popped up menu. Terms can be removed
by dragging on to the recycle bin or through the terms context menu.
We are currently implementing a number of other features, some of which
were suggested in the rst formative evaluation of this interface [2]. Among
others, we are implementing the features for obtaining the models snapshots in
the past, exporting the model, searching terms, and the undo feature.
3 Description of the Demo
For the demo, visitors will be able to assume the role of one of three news portal
users with dierent interest models, which we have created following the user
modeling approach as described in [1]. For example, one of the available interest
models is the model for Klaus, a German guy interested mostly in news about
his home country, politics, and business. As those users, visitors will be able
to use IntrospectiveViews as described above and to observe changes to a news
portal resulting from their adaptations of the interest model.
4 Conclusion and Future Work
The demo will show the application of IntrospectiveViews for scrutinizing user
interest models. The interface allows the user to get an overview of the entire
model, provides possibilities to zoom in, to lter, to obtain details including
relationships among concepts, and to change the model. In our future work, we
plan to introduce a number of features that increase the users engagement and
motivation to scrutinize, such as, e.g., a feature for comparing a persons interest
to those of communities she is a member of.
Acknowledgements. This research is carried out in the framework of the Minerva
project and is supported by IBM Deutschland Research & Development GmbH.
References
1. Bakalov, F., Konig-Ries, B., Nauerz, A., Welsch, M.: A hybrid approach to iden-
tifying user interests in web portals. In: Proc. of the 9th Int. Conf. on Innovative
Internet Community Systems (2009)
2. Bakalov, F., Konig-Ries, B., Nauerz, A., Welsch, M.: IntrospectiveViews: an inter-
face for scrutinizing semantic user models. In: Proc. of the 18th Int. Conf. on User
Modeling, Adaptation, and Personalization (2010)
3. Shneiderman, B.: The eyes have it: A task by data type taxonomy for information
visualizations. In: Proc. of the 1996 IEEE Symposium on Visual Languages (1996)
45
Demonstration: Adaptive Agent-Based Learning of
Intercultural Communication Skills
W. Lewis Johnson and Alicia Sagae

Alelo Inc., 12910 Culver Bl., Suite J, Los Angeles, CA 90066 USA
ljohnson@alelo.com, asagae@alelo.com
Abstract. This demonstration will highlight the user modeling and adaptation
capabilities of Alelos language and culture training courses. Learners acquire
intercultural competence through interactive exercises and conversations with
animated conversational agents, which play roles in immersive intercultural
communication scenarios. The conversational agents adapt their behavior based
upon the learners demonstrated skill in intercultural communication. Courses
of instruction are tailored automatically based on the learners area of interest
and learning objectives. The systems track the learners achievement of learning
objectives, to further tailor and focus the curricula.
Keywords: Pedagogical agents, language learning, curriculum adaptation
1 Introduction
This demonstration highlights learner modeling and adaptation capabilities in Alelos
interactive environments for learning intercultural communication skills. Fig. 1 shows
an example learning environment, Operational Indonesian. In this course learners can
learn the basic skills necessary to engage in overseas operations such as humanitarian
assistance. They practice their skills in interactive game scenarios. In this scenario the
learners character (center left) is engaged in a conversation with the local military
commander (center right) about providing aid. The learner communicates with non-
player characters by speaking in Indonesian into a microphone, and selecting
appropriate nonverbal gestures. The goal is for learners to engage in conversation
without hints or assistance, but until they get to that point they can refer to a list of
hints in English (top left), or in Indonesian.
Approximately 100,000 people around the world have used Alelo courses to learn
about foreign languages and cultures [1]. The courses offer excellent illustrations of
how user modeling, adaptation, and personalization can be put to practical use to
improve the effectiveness of computer-based learning.
Alelo courses make extensive use of pedagogical agents to support the learning
process. Our conversational agents fall into two main categories: conversational
partners and virtual coaches. As shown in Fig. 1, learners engage in conversation with
conversational partners during practice scenarios. Partners respond to the learners
spoken utterances and nonverbal actions, in a manner that is appropriate for the
culture, the partners social role, and the social context of the conversation. The
46
manner in which the agents respond provides learners with cues as to how well they
are performing. They may express approval when learners speak courteously, or may
express offence when the learner says something inappropriate.



Fig. 1. Operational Indonesian language and culture training system.
Agent reactions to faux pas may be subtle and easily overlooked by someone who
is not familiar with the culture, and learners sometimes do not understand what
exactly they did wrong or why it is a mistake. We therefore scaffold practice dialogs
with hints and additional feedback and explanations. Virtual coaches prepare learners
for practice scenarios and offer feedback, Auditory signals (earcons), and graphical
symbols (green plusses and red minuses) signal when the learner has done something
particularly good or bad.
4 Adaptation
Adaptation is necessary to make the behavior of the conversational partners adjust to
the level of communicative skill of the learners. Each agent has a level of rapport
with the learner, which responds dynamically to the learners use of intercultural
skills, which in turn affects how the agent responds to the learner.
It is also useful to adapt the level of difficulty of a practice scenario to the skill
level of the learner. This is currently accomplished by adjusting the amount of
additional scaffolding that is provided in the scenario. The symbols and earcons that
47
signal a change in the agents attitude and reaction, along with other assists such as
subtitles and translations, can be disabled to increase difficulty and realism.
The systems dynamically tailor the curricula to reflect the learning objectives of
each learner. At the beginning of the course the learner can specify his or her country
of interest, level of seniority and envisioned job responsibilities overseas. The system
automatically assembles a set of recommended lessons.
We are incorporating a learner model into the architecture that tracks the learners
progress in mastering the course material. This model tracks the learners progress in
each curriculum section, and dynamically selects and presents instructional materials
until the content is mastered. We have also experimented with tracking the learners
degree of mastery of each word and phrase, and detecting signs of skill decay for
particular phrases. We use this information to reinforce skills that are decaying.
In our current work we are extending our agent architecture to increase the level of
flexibility and adaptability that is supported. The new architecture, called VRP
(Virtual Role Player) [2], incorporates explicit representations of the physical and
social environment, and rules governing agent behavior. By making these
representations part of a shared state across multiple dialog instances, we can create
agents whose behavior adapts over a series of episodes to the learners
communicative competence. This results in practice experiences that are more
realistic and appropriately challenging for learners. A new project named CultureCom
is developing formal models of the cultural influences underlying dialog and using
them to increase the flexibility of the behavior of non-player characters in training
simulations. The axioms of the formal model can be loaded into an enhanced version
of VRP, allowing agent behavior to be partially conditioned on these explicit
validated models of sociocultural reasoning. The models can swapped in and out to
enable agents to model a variety of different cultural characteristics.
5 Technical Requirements
The demonstration will be performed on a combination of laptop computers and
mobile devices. We will provide our own computing equipment. Access to the
Internet for Web-based demonstrations will be desirable.

Acknowledgments. The author wishes to express his thanks to the members of the
Alelo team who contributed to this work. This work was sponsored by PM TRASYS,
Voice of America, the Office of Naval Research, and DARPA. Opinions expressed
here are those of the author and not of the sponsors or the US Government.
References
1. Johnson, W.L.: Serious Use of a Serious Game for Learning Foreign Language, IJAIED (in
press)
2. Johnson, W.L.: Using Immersive Simulations to Develop Intercultural Competence. In Proc.
of the Intl. Conf. on Culture and Computer. Springer-Verlag, Berlin (in press)
48
GDR: A Language for Flexible Interlinking,
Integrating, and Enriching User Data
Erwin Leonardi
1
, Jan Hidders
1
, Geert-Jan Houben
1
,
Fabian Abel
2
, Eelco Herder
2


1
Delft University of Technology, PO Box 5031, 2600 GA Delft, the Netherlands
{e.leonardi, a.j.h.hidders, g.j.p.m.houben}@tudelft.nl
2
L3S Research Center, Appelstrasse 9a, 30167 Hannover, Germany
{abel, herder}@l3s.de
Abstract. The Grapple Derivation Rule (GDR) language is a rule language that
is used to interlink, integrate, and enrich user data from multiple distributed
user data repositories as well as data published as Linked Data on the Web
(such as DBpedia and GeoNames). To demonstrate this, we have applied GDR
in the Grapple User Modeling Framework (GUMF) a framework that
facilitates the brokerage of user profile information and user model
representations for adaptive systems to leverage user data for the purpose of
personalization in e-learning systems.
1 Introduction
Adaptation and personalization become important features offered by todays Web
applications and services. To be able to provide such personalized and adapted
contents and services, these applications explicitly or implicitly collect data about
their users and their behavior, and thus build up user profiles. Furthermore, these Web
applications are becoming increasingly connected. This creates the interesting
challenge of performing user modeling and personalization across application
boundaries. It requires approaches allowing various Web applications to exchange,
reuse, interlink, and integrate user data. We also observe that there is a growing effort
to make data interlinked and freely available and accessible on the Web following the
principles of Linked Data
1
. This data is typically published as RDF (c.f. http://www.
w3.org/RDF) and accessible through a SPARQL (c.f. http://www.w3.org/TR/rdf-
sparql-query) endpoint. This effort opens opportunities to unlock a huge potential of
background data, to complement and enhance the user data. By reusing this
interlinked data (such as DBpedia
2
and GeoNames
3
), various relationships between
user data can now be derived and discovered, and thus make user data more
meaningful and richer. This opens opportunities for Web applications to provide

1
http://www.w3.org/DesignIssues/LinkedData .html
2
http://dbpedia.org/
3
http://geonames.org/
49
better adaptation and personalization to their users as they have more knowledge
about the users.
In this paper we demonstrate the Grapple Derivation Rule (GDR) language [1] that
allows for the integration of user data from multiple logically distributed user data
repositories and background data published as Linked Data on the Web. The
important features that GDR has are the following: (1) It provides the basis for
reasoning over distributed user data. (2) It offers a view-based query mechanism to
the applications. (3) It allows the user data to be extended and enriched using other
user data repositories and Linked Data available on the Web. (4) It gives a flexible
way of integrating, interlinking, and enriching user data. To this end, we have
integrated GDR into GUMF [2] in the context of the GRAPPLE project
4
to enable the
client applications to create flexible, rule-based plug-ins for leveraging user data that
go beyond the features of RuleML [3] or SWRL [4].
2 GDR
In human-readable syntax, a GDR rule has the form: a c, where a and c are the
antecedent and consequent of the rule, respectively, and where a is a conjunction of
premises written as p
1
... p
n
. The premises of a GDR rule are classified into two
types: dataspace premises and external source premises. A dataspace premise
describes conditions over a Grapple dataspace [2], e.g. a user data repository, in the
form of a pattern-based Grapple Query [2]. This will provide higher-level abstraction
of the condition definition in the premise. An external source premise specifies
conditions in the form of basic graph patterns over an external data source, e.g.
background knowledge, accessible through its SPARQL endpoint. As each Grapple
dataspace has its own SPARQL endpoint, it is also possible to define an external
source premise over a Grapple dataspace. The consequent describes knowledge
modeled as Grapple statements that will be derived if all the premises hold.
Let us consider a simple example of using GDR to enrich user data. Suppose an
adaptive application wants to select the language in which it presents pages to its user
based on the country where the user lives. Unfortunately, the application has only
information about the city where the user lives in her user profile. To address this, the
administrator of the application defines a GDR rule as depicted in Figure 1 over the
applications own data space and background knowledge. This rule specifies that if
the name of the city where a user lives is cityName and the city is located in a country
with name countryName, then the name of the country where the user lives is
countryName. The relation between city and country can be discovered by exploiting
knowledge available in an external data source, namely, GeoNames. The rule not only
reasons over the user data that the application has, but also extends and enriches it
with knowledge from external data sources. It also provides a view-based query
mechanism to the application, as the users country information is dynamically
computed and available from another external data source. With this rule language,
applications can easily integrate user data from other applications they have access to,

4
http://www.grapple-project.org/
50
or add external data sources, by modifying the rules without changing the
applications. Thus, it provides flexibility to the administrators of adaptive
applications.



Fig. 1 An Example of a GDR Rule in XML Syntax (partial view)
3 Demonstration Overview
Our demonstration showcases the functionality of GDR in integrating, interlinking,
and enriching user data. For this, we use the GDR implementation that has been
integrated into GUMF, as part of the research performed in the EU project
GRAPPLE. In the demonstration, we will: (1) show how GDR is used for the
integration between distributed user data, (2) demonstrate the enrichment of user data
with data published as Linked Data on the Web, (3) showcase the view-based query
mechanism offered by GDR, and (4) demonstrate the flexibility that GDR offers in
integrating, interlinking, and enriching user data.

Acknowledgments. This work was partially supported by the EU FP7 project
GRAPPLE (Generic Responsive Adaptive Personalized Learning Environment)
References
1. Leonardi, E., Abel, F., Heckmann, D., Herder, E., Hidders, J., Houben, G.J.: A Flexible
Rule-Based Method for Interlinking, Integrating, and Enriching User Data. In the Proc. of
ICWE 2010, Vienna, Austria, July, 2010.
2. Abel, F., Heckmann, D., Herder, E., Hidders, J., Houben, G.J., Krause, D., Leonardi, E.,
van der Sluijs, K.: A Framework for Flexible User Profile Mashups. In Proc. of APWEB
2.0 2009 Workshop in conjunction with UMAP 2009, Trento, Italy, June, 2009.
3. Rule Markup Language Initiative. Rule Markup Language (RuleML). http://ruleml.org/
4. Horrocks, I., Patel-Schneider, P. F., Boley, H., Tabet, S., Grosof, B., Dean, M.: SWRL: A
Semantic Web Rule Language - Combining OWL and RuleML. W3C Member
Submission, 2004.
51
A Motivational Mobile Application Toolset for Emotional
Eating
Kendra Markle, Lorin Wilde

kendra@alum.mit.edu, wildercom@gmail.com
Abstract. Mobile phone applications can be used to combat emotional eating,
emotion induced increases in consumption of food, by facilitating behavior
tracking, detecting trends and providing real time feedback. We describe the
design of a personalized mobile application with features to motivate
decoupling of food and mood behaviors. Specifically, we look at the Emotional
Eating Scale (EES) subscales most highly associated with over-eating
(angry/frustrated, anxiety, and depression), correlate them with eating trends
and apply motivational techniques, such as cognitive reappraisal and setting
intentions, in a personalized feedback system. This smart phone application
facilitates efficient tracking of moods and food eaten with voice recognition,
automatically determines calorie intake, detects key affective and caloric
consumption patterns, provides motivational feedback at appropriate times in a
variety of ways and reinforces positive behavior change.

Keywords: emotional eating, obesity, persuasive technology, mobile health
1 Introduction
Emotions can induce changes in eating [1], and many people, both obese and
otherwise, over-eat to regulate or avoid their emotions [2][3][4], especially when lack
of basic needs causes emotional stress [5]. Engaging in emotional eating is generally
not a conscious act, which makes understanding and changing one's own behavior
difficult. With two thirds of American adults being overweight or obese [6], tools that
heighten awareness of personal associations between moods and eating are of
growing importance. An overall objective of our system is to guide over-eaters to
healthier eating patterns by removing these types of associations. We use voice
recognition to facilitate self-reported tracking of both food and moods and a
motivational framework that detects key behavior trends and provides reinforcement
to facilitate breaking those patterns. Our application looks for caloric intake
correlations with emotional patterns described by the top three emotion subscales that
correlate with overeating [7]: angry/frustrated (angry, inadequate, frustrated), anxiety
(guilty, bored, anxious), and depression (lonely, tired, depressed).
Our demonstration employs a combination of motivational techniques described
below, which have been compiled from the literature and from both empirical and
anecdotal evidence. We also draw from relevant literature on longitudinal health
tracking [8] and characteristics of effective just-in-time information delivery [9].
52
2 Demonstration: Motivational Mobile Application Toolset
When emotional eating patterns are detected in the tracked data, the system
provides personalized feedback to build awareness and motivate behavior change in
the form of instant popup messages and timed notifications.

Assignments & Positive Reinforcement: After detecting a behavior pattern, the
application will give behavior assignments to the user to motivate them to take small
steps to change. For example, the user might receive a pop up message saying "I've
noticed over the last few days that you're eating while you're anxious. This is
probably contributing to your over-eating. Can you think of some ways to reduce your
stress? [Yes|Need Suggestion]". If the user answers "Yes", they get a choice of
assignments. If they request a suggestion, an assignment is given to them. The
application then evaluates their progress on the assignment over the next few days. If
the target behavior is seen (in this case, less anxiety reported with meals), then the
user will get a randomly timed popup message with positive reinforcement, such as
"Something's working really well youre reporting less anxiety at mealtimes. Keep
an eye on this and keep up the great work".

Cognitive Reappraisal of Progress: Change is a cyclical process where we learn
by making mistakes. Teaching someone to reappraise their initial reaction to a setback
or disappointment is imperative for helping them get through the hard times when
making change. Our application helps users cognitively reappraise their situation
when it detects a setback. For example, a message might say "Even during this tough
week, you're making progress in learning what to look for and that conscious
awareness will translate into action when you're ready".

Intentions and Principled Stances: In situations where repeated failure has
already undermined self-efficacy, focusing on intentions that are tied to principled
stances, deeper held beliefs about identity, can help reframe progress in the positive
and reduce the sense of failure. Small steps towards making change are rewarded and
the overall experience is a series of small successes instead of failures. For example,
instead of using goal language such as I will not eat any dessert this week, a
suggested commitment would focus on the intention and connect it to a deeper belief
the user has, such as Today, I will try hard to eat less sugar than I have recently,
because that much sugar is not good for me and my health is really important.

Framing Behavior Implications: When feeding back information about behavior
trends, some messages are written with context and/or projections about the future to
impress upon the user the effects of their behavior on their health. For example, a
message might say "Did you miss lunch again today? Skipping lunch repeatedly
stresses your body and puts it into weight gain mode. After a few weeks of this, you're
likely to see more fat on your body".

Authority, Partnership & Trust: Our messages, timing and tone are all designed
to build trust, so that we can alternately use an authoritative voice and work in
partnership with the user on a specific goal. For example, after detecting a targeted
53
behavior, a message might say "You've done this a number of times - I want you to
stop" (authoritative) or "This might feel discouraging, but we will find a solution
together" (partnership).

Buddy System Support: The application allows users to specify buddies they
know who will support them when they need to talk to a friend. The application
identifies times when reaching out to a buddy could help, such as when they indicate
they're lonely for several days in a row, and acts as a speed dial to connect the user
with their buddy.

Crowd-sourced Motivational Messages: After positive behavior change is seen,
the application asks the user what was most helpful and what message they wish they
had received, knowing what they know now. The messages they write can then be
used with other users struggling with the same behavior.
3 Future Work
While incorporating new features to make this application a more effective and
engaging tool for fighting obesity, we will endeavor to improve the user interface
timing and make it more affective. Because emotional eating isn't a rational problem,
rational communication alone is unlikely to cause significant and lasting change and
the affective capability of empathizing with the user's current emotional state and
using an appropriate message tone will be an important enhancement.
References
1. Macht, M.: How Emotions Affect Eating: a Five-Way Model. In: Appetite, 2008 vol. 50,
no. 1, p1-11
2. Ganley, R.: Emotion and Eating in Obesity: A Review of the Literature. In: International
Journal of Eating Disorders, 2006 vol. 8, no. 3, p343-361
3. Waller, G., Osman, S.: Emotional Eating and Eating Psychopathology Among Non-eating-
disordered Women. In: International Journal of Eating Disorders, 1998 vol. 23, no. 4,
p419-424
4. Spoor, S., Bekker, M., Strien, T., van Heck, G.: Relations Between Negative Affect,
Coping and Emotional Eating. In: Appetite, 2007, vol. 48, no. 3, p368-376
5. Timmerman, G., Acton, G.: The Relationship Between Basic Needs Satisfaction and
Emotional Eating. In: Issues in Mental Health Nursing, 2001 vol. 22, no. 7, p691-701
6. The Obesity Society, http://www.obesity.org/statistics
7. Arnow, B., Kenardy, J. and Agras, W.S.: The Emotional Eating Scale: The Development
of a Measure to Assess Coping with Negative Affect by Eating. In: International Journal
of Eating Disorders, 1995 vol. 18, no. 1, p79-90
8. Beaudin, J., Intille, S. and Morris, M.: To Track or Not to Track: User Reactions to
Concepts in Longitudinal Health Monitoring. In J. Med Internet Res, 2006 vol. 8, no. 4. 1-
18
9. Rhodes, B.: Just-In-Time Information Retrieval. Ph.D. Dissertation, MIT Media Lab, May
2000
54
Learning by Teaching SimStudent
Noboru Matsuda
1
, Kenneth Koedinger
1
, William Cohen
1
,
Victoria Keiser
1
, Rohan Raizada
1
, and Gabriel Stylianides
2


1
Carnegie Mellon University
2
University of Pittsburgh
5000 Forbes Ave. Pittsburgh PA 15213 5517 Posvar Hall, Pittsburgh PA 15260
Noboru.Matsuda@cs.cmu.edu
Abstract: We propose a live demo of our latest on-line game-like learning en-
vironment where students learn to solve algebra equations by teaching a teach-
able computer agent, called SimStudent.


The purpose of our proposed demo is to show how our cutting edge technology of
a teachable agent, called SimStudent, can be used for students to practice learning by
teaching. It is well known that students learn by teaching others [1].
Despite ample empirical evidence that supports such an effect of tutor learning, we
have yet to investigate underlying cognitive and social principles that govern tutor
learning. Furthermore, the tutor learns by teaching at the cost of the tutee the tutee
might not learn much especially when the tutor himself is learning the subject to tutor.
To address these issues advancing the theories of the tutor-learning effect without
interfering with tutees learning, we have developed an on-line game-like learning
environment where students learn by teaching SimStudent. Although our Learning by
Teaching environment (LBT environment, for short) is domain independent, we will
use algebra linear equations for an example subject domain as shown in Figure 1.
SimStudent is a machine-learning agent that inductively learns domain principles
from examples [2]. In the context of learning by tutoring, a student plays the role of a
tutor and interactively tutors SimStudent. Namely, the student poses a problem for
SimStudent to solve. SimStudent attempts to solve the problem. The student then
provides feedback for each step performed by SimStudent indicating the correctness
of the attempt made by SimStudent (i.e., so-called flagged feedback). If the student
provides negative feedback to the SimStudent attempt, SimStudent may make another
attempt until either the step is eventually performed correctly or SimStudent cant
make an alternative attempt. If SimStudent cannot perform the step correctly, SimS-
tudent asks the student for a hint on what to do next. The student then provides a
hint by demonstrating the step for SimStudent. SimStudent accumulates examples
of skill applications (1) when the student demonstrates steps as hints (positive exam-
ples), and (2) when the student provides feedback for the steps performed by SimStu-
dent (both positive and negative examples).
55
Figure 1. A screenshot of the Learning by Teaching Environment. SimStudent is visualized at
the bottom left, called Lucy. The [Example n] buttons show worked-out examples for the stu-
dents to review how to solve equations. The goal of the student is to tutor Lucy well enough so
that Lucy passes the quiz, which is available at the bottom of the interface (i.e., the [Quiz Lucy]
button).

The above tutoring context provides a naturalistic dialogue between the student
and SimStudent. The student must diagnose the SimStudent proficiency level by
observing the SimStudents performance, i.e., the correctness of the steps performed.
SimStudent also models human students misconceptions [3]. At the beginning,
SimStudent is pre-trained to solve simple equations (e.g., one-step equations), but also
makes some of the errors that human students typically make. The student must use
56


their observations of SimStudents behavior to identify the potential causes of such
errors and then provide appropriate problems to remediate these misconceptions.
In the LBT environment, the student is informed that the goal is to tutor SimStu-
dent well enough so that SimStudent passes the quiz, which is prepared by the system
developer. At the bottom of the screenshot (Figure 1), there is a [Quiz Lucy] button
to have SimStudent take a quiz. The summary of the quiz results appears in a separate
window for the student to review.
Since the student is learning equation solving, he/she might not be able to tutor
SimStudent correctly. Especially when SimStudent asks for a hint on what to do next,
the student might not be able to demonstrate how to perform the next step. There are a
number of worked-out examples available for the student to review. The [Example n]
tabs at the top of the screen provide worked-out examples. When the student clicks
on one of these example buttons, a worked-out example appears on the tutoring inter-
face, just like the one shown in Figure 1 with additional explanations in the text box
below the tutoring interface.
References

1. Roscoe, R.D., Chi, M.T.H.: Understanding tutor learning: Knowledge-building and
knowledge-telling in peer tutors' explanations and questions. Review of Educational
Research. 77(4), 534--574 (2007)
2. Matsuda, N., Cohen, W.W., Koedinger, K.R.: Applying Programming by Demonstration in
an Intelligent Authoring Tool for Cognitive Tutors. In: AAAI Workshop on Human
Comprehensible Machine Learning (Technical Report WS-05-04), pp. 1--8. AAAI
association, Menlo Park, CA (2005)
3. Matsuda, N., et al.: A Computational Model of How Learner Errors Arise from Weak Prior
Knowledge. In: Taatgen, N., van Rijn, H. (eds.) Proceedings of the Annual Conference of
the Cognitive Science Society 2009, pp. 1288--1293. Cognitive Science Society: Austin,
TX (2009)
57
MARY: A Personalised Virtual Health Trainer
Hien Nguyen, Judith Masthoff


Computing Science Department, University of Aberdeen
{n.nguyen, j.masthoff}@abdn.ac.uk
Abstract. This paper describes MARY, a personalised virtual health trainer that
promotes regular walking to inactive people. In particular, we discuss the
systems theoretical basis, behavioural techniques being used, mode of delivery,
and how the content is delivered.
1 Introduction
This paper describes MARY, a personalised virtual health trainer that promotes regular
walking to inactive people. In Section 2, we present the design of the system. In
Section 3, we briefly discuss the results of an evaluation study and our future plans.
2 MARYs design
2.1 Theoretical basis
The Theory of Planned Behaviour suggests that Intention is the direct determinant of
any volitional behaviour. In turn, Intention is determined by Attitude towards the
behaviour, Subjective Norm - the perceived views of important others, and Perceived
Behavioural Control - the extent to which we feel that the behaviour is easy to
perform and/or under our control [1].
In reality, there is often a gap between our intention and our behaviour. Self-
regulation Theory states that success in implementing our intention requires self-
regulation. First, we must actively monitor our behaviour, and evaluate how this
behaviour affects our personal goals. If it hinders our goals, we drop the behaviour.
Vice versa, we reinforce the effect by adopting or continuing the behaviour [2].
Operant Learning Theory states that we change our behaviour through rewards and
punishment. We adopt a new behaviour if it leads to favourable outcomes ((e.g.
verbal praise, a good grade, a feeling of satisfaction) or removes unfavourable
outcomes and vice versa [3].
Finally, Relapse Prevention suggests various strategies to identify and prevent
high-risk situations where we often fall back to our old behaviour [4].
58
2.2 Behavioural change techniques and mode of delivery
The theoretical basis enables us to target the right determinants of behaviour and to
identify the techniques that can be used to influence such determinants. MARY
implements and personalises 12 behavioural change techniques: give general
information, prompt intention formation, identify barriers, set specific goals, self-
monitor, give feedback, review goals, use prompts, use rewards, action plans, coping
plans, and motivational interviewing.
The content is then delivered through a web-based interface, email or text-
messaging service, which allows both the users and MARY to initiate contact with
each other. This differentiates MARY from the majority of other systems where
interactions require the users initiatives. Such a one-way relationship would normally
cause the proactive party (the users) to eventually let go of the relationship. A mixed-
initiative model could make the system more natural, hence enhance its ability to
build and prolong the relationship with its users [5].
To enhance MARYs credibility, we have personified MARY as a sport instructor
using a photo of a real person. The photo was validated in an independent study [4].
We have also empirically studied the positive impact of the motivational interviewing
approach [6] and the potential for Mary to provide emotional support [7].
2.3 The intervention
MARY delivers the intervention through a series of meetings with the users. A
chronological outline of the meetings is shown in Figure 1.


Figure 1. The intervention flow

During the initial meeting, MARY attempts to establish the users confidence and trust
in the system, assesses their attitudes towards regular walking, and informs them the
importance of self-monitoring and active participation. At the end, the users are asked
to monitor their daily walks using a pedometer. After a week, an email reminder
containing a link to the first weekly meeting is sent to the users.
In the first weekly meeting, MARY follows up on the initial meeting, gives
personalised feedback on the users baseline step count, and encourages them to set a
new weekly goal (increase their step count by 10 or 20%). MARY also encourages
them to have a regular weekly meeting with her. MARY always stresses the active role
of the users during these meetings (e.g. the user decides which topics to talk about).
In a daily meeting, MARY might start with some social dialogues (e.g. Im so glad
to see you again. Its been such a long time. I thought you might have lost interest in
me.). Next, the users are asked to enter their pedometer readings. In return, MARY
Initial
meeting
First weekly
meeting
Daily
meeting
Weekly
meeting
10 minutes 25 minutes 15 minutes

Daily
meeting
2-3 minutes 2-3 minutes
59
gives a personalised feedback based on their goal, and past performance (e.g. Youve
reached your goal 3 times this week. You should treat yourself to something nice).
The daily interaction is kept to a minimum to reduce the users time and effort.
MARY also carries out a weekly meeting to review and discuss the users progress,
their weekly goal, difficulties encountered, and solutions.
The users are also provided with an interactive diary (see Figure 2, left). The diary
allows them to keep track of their progress and rewards (bronze, silver, or gold
medals), set up action plans and reminders of their plans via email and/or text
messages (e.g. Enjoy your walk to the park with the kids this morning.).




Figure 2. Screenshots of the users online diary and MARY giving feedback
3 Conclusions and Future Work
So far, we have evaluated MARY with 25 users who used MARY for at least 4 weeks.
The results were remarkably positive. An improved MARY is being developed based
on the users feedback. Notably changes include a recommendation engine that
converts calories to personalised food items, consideration of negative feedback, and
visualisation of step count as personalised routes. The new version of Mary will be
evaluated using the MRC framework for evaluating complex intervention [8].
References
1. Ajzen, I. (1991). The Theory of Planned Behavior. Organizational Behavior and Human
Decision Processes, 50, 179-211.
2. Carver, C., and Scheier, M. (1981). Attention and Self-Regulation: A Control-Theory
Approach to Human Behavior, Springer.
3. Skinner, B.F. (1953). Science and Human Behavior.
4. Nguyen, H. and Masthoff, J. (2007). Is it me or is it what I say? Source image and
persuasion. Persuasive Technology Conference. California, CA, USA, 231-242.
5. Campbell, R., Grimshaw, M., Green, G. (2009). Relational agents: A critical review. The
Open Virtual Reality Journal, 1, 1-7.
6. Nguyen, H., Masthoff, J. (2008). Designing persuasive dialogue systems: Using
argumentation with care. Persuasive Technology conference, Oulu, Finland, 201-212
7. Nguyen, H., Masthoff, J. (2009). Designing empathic computers: the effect of multimodal
empathic feedback using animated agent. Persuasive Technology conf., Claremont, USA
8. Developing and evaluating complex interventions: new guidance.
http://www.mrc.ac.uk/complexinterventionsguidance
60
A Generic Privacy-Enhancing Personalization
Infrastructure
Yang Wang and Alfred Kobsa
Donald Bren School of Information and Computer Sciences
University of California, Irvine, U.S.A.
{yangwang,kobsa}@uci.edu
Abstract. Respecting users privacy is a key challenge for building personalized
systems. We have designed a generic privacy-enhancing personalization infras-
tructure that allows system designers to express privacy constraints at design time
and enables system enforcement of privacy constraints at runtime.
1 Introduction
As personalization emerges as a mainstream practice for web sites, privacy issues only
get heightened. A recent example is Facebook Open Graph. Open Graph is an inno-
vative technology that gives its users instant personalized experience in Facebook
partner sites (e.g., cnn.com and yelp.com) without users logging in these sites [1]. De-
spite its appeal, it also ratchets up privacy concerns [2]. To reconcile privacy and
web personalization, two main types of privacy constraints need to be taken into con-
sideration in designing and implementing web-based personalized systems: regulatory
privacy requirements set out by various privacy laws and regulations [3], and users
personal privacy preferences/needs [4].
2 Infrastructure Functionalities
We have designed a generic privacy-enhancing personalization infrastructure that: (1)
supports system designers to graphically express privacy constraints, their interdepen-
dencies as well as their impacts on the personalized system at design time, (2) allows
end users to set their personal privacy preferences and to see how the personalized sys-
tem operates in accordance with their current privacy settings, and (3) enables run-time
dynamic enforcement of the privacy constraints in the system.
To our best knowledge, all three features above are novel in personalized systems.
See [5] for a more comprehensive review of related work in this area.
2.1 Support System Designers to Express Privacy Constraints
System designers of privacy-enhancing personalized systems need to capture two do-
main models: (1) a systemmodel representing the systemfeatures/components and their
interdependencies, and (2) a privacy model expressing the privacy constraints, their in-
terdependencies, and the impacts of these privacy constraints on the system features.
61
We use the notions of change sets and relationships as the underlying constructs to
represent the two domain models [6]. A change set is an individually-identiable set
of changes from a base version of a system. A particular version of the system can be
composed by mixing and matching different change sets. Relationships are constraints
that govern the system composition from change sets. Relationships among change sets
are constructed using logic operators such as AND and OR. We design four types of
change sets in our modeling tool: feature change sets and relationships which are used
to model the software model, preference change sets and country change sets which are
used to model the privacy model, and mapping change sets which are used to connect
the two models.
Fig. 1. Modeling System Domain, Privacy Domain, and Their Interdependencies
Figure 1 illustrates the usage of our modeling tool for a personalized system. Each
row represents a change set and each column to the right of the Change Set column
represents a relationship. The two domain models remain independent, interconnected
only through the common mapping change sets. This approach makes the two models
less coupled and thus easier to model and manage. This tool also makes it straightfor-
ward to track which country and what specic privacy constraints have been supported.
This would make both internal and external audits easier.
2.2 Support Users to set Privacy Preferences and See Consequences Instantly
This infrastructure also enables users to express their privacy preferences with regard to
a personalized system. More importantly, users can see immediately the consequences
of their privacy changes/settings in terms of how the underlying personalized system
might change (e.g., certain system components get turn on or off). Figure 2 shows an
example of this functionality. The upper part contains the privacy options that a user can
specify, while the lower part lists the usage of personalization components according to
the current privacy settings.
62
Fig. 2. Privacy Control Tool for Users
2.3 Enforcement of Privacy Constraints at Runtime
Internally, each personalization component is associated with a Boolean guard. The
current privacy settings are translated as values that are used to evaluate these Boolean
guards. A personalization method would be turned on if its Boolean guard is evaluated
to be true, and vice versa [7].
References
1. Facebook: Open graph protocol. http://developers.facebook.com/docs/opengraph (2010)
2. Shuler, P.J.: Facebooks open graph ratchets up privacy concerns.
http://www.npr.org/templates/story/story.php?storyId=126183577 (2010)
3. Wang, Y., Chen, Z., Kobsa, A.: A Collection and Systematization of International Privacy
Laws, with Special Consideration of Internationally Operating Personalized Websites. (2006)
4. Teltzrow, M., Kobsa, A.: Impacts of user privacy preferences on personalized systems: a com-
parative study. In: Designing personalized user experiences in eCommerce. Kluwer Academic
Publishers (2004) 315332
5. Wang, Y., Kobsa, A.: Technical solutions for Privacy-Enhanced personalization. In: Intelli-
gent User Interfaces: Adaptation and Personalization Systems and Technologies. IGI Global,
Hershey, PA (2009) 353376
6. Wang, Y., Hendrickson, S.A., van der Hoek, A., Taylor, R.N.: Modeling PLA variation of
Privacy-Enhancing personalized systems, San Francisco, CA (2009) 7180
7. Wang, Y., Kobsa, A. In: Respecting Users Individual Privacy Constraints in Web Personal-
ization. Berlin - Heidelberg - New York: Springer-Verlag, Corfu, Greece (2007)
63
SocConnect: Intelligent Social Networks Aggregator
Yuan Wang
1
, Jie Zhang
2
, and Julita Vassileva
1
1
Department of Computer Science, University of Saskatchewan, Canada
2
School of Computer Engineering, Nanyang Technological University, Singapore
1
{yuw193, jiv}@cs.usask.ca
2
zhangj@ntu.edu.sg
Abstract. We have developed a dashboard application called SocConnect for
integrating social data from different social networking sites (e.g. Facebook,
Twitter), which allows users to create personalized social and semantic contexts
for their social data. Users can blend their friends across different social
networking sites and group them in different ways. They can also rate friends
and/or their activities as favourite, neutral or disliked. Machine learning can be
usefully applied in predicting the interest level of users in their social network
activities, thus helping them deal with cognitive overload.
1 Introduction
Social Networking Sites (SNSs) have changed how people communicate: nowadays,
people spend more time on SNSs than ever, and prefer communication via SNSs over
emails. Current SNSs have the limitation of poor user data interoperability [1]. User
content, online activities, and friends are scattered over different places. It becomes
increasingly inconvenient for users to manage their social data and constantly check
many sites to keep track of all recent updates. People may also keep different
accounts on the same SNS in order to protect their privacy or other purposes. In
addition, users are often overwhelmed by the huge amount of social data, especially
friends activities. There have been many attempts to create social network
aggregators that integrate a users accounts on different SNSs. Based on their
platforms, social aggregators can be classified as web-based and desktop applications.
In web-based aggregators, users need to register and create a new account for the
aggregator, and provide their SNSs accounts information to the aggregator. In desktop
aggregators, which have been emerging on mobile platforms recently, users do not
need to create an account. Based on their functions, social aggregators can be divided
into three groups: write-only, read-only, and write and read. Write-only and read-only
aggregators usually are lightweight and web-based. They allow users to publish the
same status update to multiple SNSs. However, these applications do not allow users
to blend or group their friends from different places; therefore the updates appear
scattered, out of context.
In contrast, we take a read-only desktop-based approach for integrating the
users social data on different SNSs, and that allows users to organize and to create
personalized contexts for their social data. We also provide personalized
64
recommendation of friends activities from different SNSs. While many SNSs deploy
algorithms based on the analysis of social network structure to recommend new
friends to the user, there havent been many approaches to recommend contents on
SNSs. One such approach is SoNARS, which combines results from collaborative
filtering and content-based recommendation [2]. A Twitter desktop client application
called TalkingPuffin (talkingpuffin.org) allows users to remove noise (uninteresting
updates) by manually muting users, retweets from specific users or certain
applications. Our approach provides content-based recommendations by applying
machine learning techniques on previous rated by the user social activities.
2 SocConnect Dashboard
The SocConnect dashboard [3] is a client application communicating with a server
that processes the data and generates recommendations. It retrieves the information
about the users friends and their activities (streams) using the APIs of different
SNSs. SocConnect provides three functional categories, managing friends, rating
friends and activities, and recommendation of activities.
The first functional category, managing friends contains two functions:
blending friends and grouping friends. Blending friends allows the user to merge the
different accounts of a given friend across two or more SNSs, and to create a single
blended friend account for this friend in the dashboard. The second function is to
group friends. Users can put their friends, both individual SNS accounts and
integrated accounts into named groups. This function allows users to express the
context and semantics of friendships, which could be the shared characteristics,
interests or activities between friends. The second functional category, rating friends
and activities, allows users to rate friends or friends activities as favourite or disliked.
The favourite activities are bookmarked, and be revisited easily.
The third functional category, personal recommendations automatically
provides recommendation of activities that may be interesting based on the previous
ratings and the information about friend groups. A content-based approach is used to
predict possible favourite or disliked activities by the user. The activities that were
rated by the user in the past are fed to machine learning algorithm which is trained to
predict the level of liking that the user may have of an incoming activity. Both textual
(vector space) and non-textual features of the activities are used for learning to predict
which activities will be found interesting by the user. The following non-textual are
used: the actor (the friend who originated the activity), the actor type (if the friend is
marked as favourite, or disliked), the activity type (upload photo album, share link,
upload a photo, upload a video, status update, application use, reply, retweet), the
SNS that is the source of the activity (Twitter of Facebook), the application that is the
source of the activity (e.g. Farmville or ForSquare), and user rating of the activity.
Based on these features the machine learning algorithm is trained to predict the level
of liking of future incoming activities, which are classified into 5 possible levels
varying from strong dislike to strong like. The incoming activities in the combined
65
stream of the integrated SNSs is displayed in a way that highlights the recommended
activities (see [3] for details).
We experimented with four different machine learning algorithms in WEKA to
see which one will generate the best predictions. These algorithms (methods) were:
Decision Tree, Radial Basis Function (RBF), Naive Bayesian Network and SVM
(Support Vector Machine). We compared also the quality of recommendations using
only textual data, only non-textual data and a combination of both textual and non-
textual data. We found that when using only non-textual features, although the
performance difference among the algorithms is not significant, support vector
machine (SVM) provides the best performance, and it correctly classifies 69.9% of
instances in the testing data. Using only the textual features, the performance is not
very good (72% for BayesNet, 75% for SVM, 77% for Decision Tree, and 84.5 for
RBF). Combining textual and non-textual features yields the best performance overall
(75% for BayesNet, 80% for SVM and Decision Tree, 81.4% for RBF). Therefore,
even though more computationally expensive, using the text features is worthwhile.
3 Conclusions and Future Work
We have created a novel desktop based SNSs integrator, called SocConnect. It
allows users to blend their friends across different SNSs, to group them and to rate
their friends and activities. This integrator is used as a platform for generating
personalized recommendations of activities for the user, to help users deal with the
information overload. We have investigated different machine learning methods to
learn about implicit user preferences and generate predictions and found one, RBF
that looks particularly promising. In our future work we will explore more deeply the
importance of different features of SNS activities. We will conduct user studies with
the SocConnect prototype to test its usability. We will also work on expanding its
functionality to read and write rather than just read.
4 References
1. F. Carmagnola, F. Vernero, and P. Grillo, Sonars: A social networks-based algorithm for
social recommender systems, in Proc. 17
th
International Conference on User Modeling,
Adaptation, and Personalization, 2009.
2. G. Ereteo, M. Buffa, F. Gandon, M. Leitzelman, and F. Limpens, Leveraging social data
with semantics, in Proc. of the W3C Workshop on the Future of Social Networking, 2009.
3. Y. Wang, J. Zhang, and J. Vassileva, SocConnect: A user-centric approach for social
networking sites integration, in Proc. International Conference on Intelligent User
Interface (IUI) Workshop on User Data Interoperability in the Social Web, 2010.
4. Webster, A., Vassileva, J.: Personal relations in online communities. in Proc. Adaptive
Hypermedia and Adaptive Web-Based Systems, Dublin, Ireland, Springer LNCS (2006), 223
-233.
66