Professional Documents
Culture Documents
Social Rec Om Mender System
Social Rec Om Mender System
System
15.08.08
Agenda
Introduction
Background on Collaborative Filtering
Collaborative Filtering Limitation
Using trust in RS
Related works
Research methodology
Evaluation and Results
Conclusion
Introduction
Recommender system (RS) help users find items
(e.g., news items, movies) that meet their specific
needs.
Motivation
Information overload
Researches in RS focused on developing methods
and approaches dealing with the Information overload
problem.
Main Approaches
Content-Based (Salton, 1989)
Collaborative filtering/Social Filtering (Goldberg, 1992 )
hybrid
Collaborative Filtering (CF)
In the real world we seek advices from our trusted
people
CF automate the process of “word-of-mouth”
General use:
Weight all users with respect to similarity with the active
user.
Select a subset of the users (neighbors) to use as
predictors (recommenders).
Rating prediction:
n
w a ,u ( ru ,i ru )
pa ,i ra u 1
n
w
u 1
a ,u
User-User Collaborative Filtering
Active user
3
?
Rating
prediction
CF Limitation
New item problem
Cold start problem
Sparsity (95%-99%)
Controversial user
Easy to attacks
Scalability
Cannot recommend items to someone with
unique tastes.
Tends to recommend popular items
Solution: using trust relationships
Implicit: Deriving trust score directly from the rating
data
Generally based on user prediction accuracy in the past
Explicit: users explicitly “rate” other users
FilmTrust (Hendler et al,2006)
Molskiing (Massa et al,2005)
Limitation:
Users have on average very few links (trusted sources)
More User’s effort
Solution
Trust propagation: find unknown user’s
trustworthiness based on the users’ “web of trust”
Trust inference
Global metrics: computes a single global trust
value for every single user (reputation) 1 b 5
Examples: a 3 d
PageRank (Page et al, 1998),eBuy 2 c 3
Pros:
Based on the whole community opinion
Simple to compute
Cons:
Trust is subjective (controversial users)
Local trust metrics
Local metrics: predicts (different) trust scores that are
personalized from the point of view of every single
user
Example:
MoleTrust (Massa et al,2006)
TidalTrust (Golbeck et al,2005)
Pros:
1 b 5
More accurate
? d
Attack resistance a
Cons: 2 c 3
Ignoring the “wisdom of the crowd”
More complicated
Related works(1):Massa et al(2006)
Crawling Epinion.com
users can review items and also assign them numeric
ratings in the range 1 to 5.
Users can also express their “Web of Trust” and their
Black list
Dataset:
~50K users,~140K items,~665K reviews
487K binary trust statement
Sparsity=99.99135%
Above 50% are cold start users (less than 5 review)
Recommendation method
Using MoleTrust metric
Input
Estimated trust output
userXuser Predicted Ratings
Rating
MXN
predictor
Rating
MXN
w a ,u ( ru ,i ru )
pa ,i ra u 1
n
w
u 1
a ,u
Evaluation and results
Related works(2):Golbeck et al(2006)
FilmTrust: Online Recommender System
Users can rate films, write reviews, and express trust
statements in other users based on how much they trust
their friends about movies ratings
Rating scale from half start to four start
Trust scale from 1 to 10
Dataset:
500 users, 100 popular movies, 11,250 rating
350 users with social connection
Sparsity=77%
Recommendation method
Weight ratings by trust value
Search recursively for trusted sources
Using TidalTrust metric for trust inference
Simple Prediction method
Example:
t r
si im
Evaluation and results
Benchmarks: Pure CF and simple average
80% training and 20% testing
Using MAE metric
First analysis, using trust didn’t appear to be effective
Above 50% of the rating were within the range of
the mean +/- half star
Trust-based significantly useful only to user who
disagree with the average
Result
Limitations
Do not distinguish between various types of social
relationships
Independent Variables
(recommendation methods)
Baseline:User-based CF
(Pearson Correlation)
Hybrids method (Similarity
and Social relations) Dependent Variables
(Combination schemes) (Performance)
Social Restriction method
MAE
(Pearson Correlation)
Recipient Sources
System’s Receiver-
Source Similarity
Calculation Systems System’s Prediction
Prediction (Recommendation)
Component
Recipient-Source
similarity
(Recipient’s)
Trust, Sources’
Friendship Qualification
s
Interaction
duration, System’s Source
frequency Qualification
Component
Reputation
Prediction method 1
Hybrid method
Social relations combined with similarity (Pearson
Correlation)
Tuning the source’s weight according to his group
Group P: sources similar to the active user
Group S: sources belong to the social network of
the active user
Pa ,u .......................if (u P) P
S
S a ,u .......................if (u S )
Wa , u
f ( Pa ,u , S a ,u )..........if (u P S )
0............................Otherwise
Prediction method 2
Social restriction
Social relations used for restriction
Consider only sources belong to both groups
S and P
Using the source’s similarity
S P
Wa , u Pa, u.....if (u P S )
Simulation System Architecture
Configuration Utility
Front-end
Similarity-Based CF Hybrid CF
Social restriction CF
(Pearson Correlation) (Pearson Correlation
and Social ties)
Offline-Online boundary
User-User Similarity
generation
0.743582
0.8 0.78
0.743 80
Coverage
0.78
0.76
MAE
coverage Precision
MAE 0.742578
MAE
MAE Recall 0.76
0.742 0.74
76
0.74
0.7415
74 0.72
0.7410.72
72 0.7
0.7405
0.7
d=6 d=5 d=4 d=3 d=2 d=1 pure CF 70 0.68
0.74
100d=690 d=5 70 d=5
d=6
80 d=4 d=340method
50d=4
60 Prediction d=2 20 d=2
d=3
30 d=1
10 0pure
d=1 CF pure CF
Prediction
Social method method
Prediction
tie weight
Hybrid method: Cold start users
Impact of shared-interest sources
1.4
1.2
0.8
MAE-CF
MAE
MAE-WAA1
0.6
0.4
0.2
0
25 23 21 19 17 15 13 11 9 7 5 3 1
Number of sources
Impact of different social measures
Socialrestriction
Social restriction
Social restrictin coverage
0.84
0.82
80
0.8
0.82
70
0.78
60
0.8
0.76
50
Precision 0.78
Coverage
0.74
MAE
MAE
Recall
coverage 40
0.76
0.72
30
0.7
0.74
20
0.68
0.72
10
0.66
d=6 d=5 d=4 d=3 d=2 d=1 0
pure CF 0.7
d=6 d=6 d=5 d=5 d=4 d=3
d=4 Prediction
d=3 method d=2
d=2 d=1
d=1 pureCF
pure CF
Prediction method
Prediction method
Social restriction: cold start users
2.5
2
MAE-CF 1.5
MAE
MAE-WAA1 1
0.5
0
25 23 21 19 17 15 13 11 9 7 5 3 1
Number of sources
Conclusion
Social relationships is effective in alleviating
CF weaknesses:
Cold start problem (Social weighting and
social restriction)
Scalability problem (Social restriction)
Spammers attacks (Social weighting and
social restriction)
References
Shardanand, U., Maes, P.: Social Information Filtering: Algorithms for Automating ’Word of
Mouth’. In: Proceedings of Human Factors in Computing Systems, pp.10–217 (1995)
Herlocker, J., Konstan, J.A., Terveen, L., Riedl, J.: Evaluating Collaborative Filtering
Recommender Systems. ACM Transactions on Information Systems 22, 5–53(2004)
Massa, P., Avesani, P.: Trust-Aware Collaborative Filtering for Recommender Systems.In:
Proceedings of the International Conference on Cooperative Information Systems (CoopIS),
Agia Napa, Cyprus, pp. 492–508 (2004)1060 C.-S. Hwang and Y.-P. Chen
Avesani, P., Massa, P., Tiella, R.: Moleskiing: A Trust-Aware Decentralized Recommender
System. In: Proceedings of the First Workshop on Friend of a FriendSocial Networking and
the Semantic Web, Galway, Ireland (2004)