Download as pdf or txt
Download as pdf or txt
You are on page 1of 8

Emotion Detection through Smartphone’s Accelerometer and

Gyroscope Sensors
Orestis Piskioulis Katerina Tzafilkou Anastasios Economides
School of Science and Technology, School of Science and Technology, SMILE lab (smile.uom.gr),
International Hellenic University, International Hellenic University, Information Systems IPPS, University
Thessaloniki, Greece Thessaloniki, Greece of Macedonia, Thessaloniki, Greece
piskioulisorestis@gmail.com katerinatzaf@gmail.com economid@uom.gr

ABSTRACT 1 INTRODUCTION
Emotion recognition is essential for assessing human emotional Emotion recognition is essential for assessing human emotional
states and predicting user behavior to provide appropriate and states and predicting user behavior to provide appropriate and
personalized feedback. The wide range of Smartphones with ac- personalized feedback [1]. The automatic recognition of emotions
celerometers, microphones, GPSs, gyroscopes, and more motivate is an interesting but difficult problem and various researchers tried
researchers to explore the automatic emotion detection through to solve. For example, it is interesting to find out what types of
Smartphone sensors. To this end, mobile sensing can facilitate the music and movies cause what emotions to people. In addition, robots
data retrieval process in a non-intrusive way without disturbing will soon be used in various areas of our lives. The robots will be
the user’s experience. This study seeks to contribute to the field of more intelligent and humanized if they can ’feel the mood’ of the
non-intrusive mobile sensing for emotion recognition by detecting person with whom they interact. It is also important to recognize
user emotions via accelerometer and gyroscope sensors in Smart- the people’s emotional state and its development to determine their
phones. A prototype gaming app was designed and a sensor log app psychological health and mental welfare [2], [3].
for Android OS was used to monitor the users’ sensor data while Smartphones are now commonly used in industry for business,
interacting with the game. The recorded data from 40 users was social and entertainment purposes in people’s everyday life. To-
processed and used to train different classifiers for two emotions: a day, Smartphone devices embed several sensors, like microphone,
positive (enjoyment) and a negative (frustration) one. The valida- accelerometer, electronic compass, GPS, proximity, etc. The data
tion study demonstrates a high prediction of 87.90% for enjoyment obtained from the sensors has profound information and can be
and 89.45% for frustration. Our findings indicate that by analyzing used to evaluate the social habits of the user, such as physical ac-
accelerometer and gyroscope data, it is possible to make efficient tivity, social communications, location, etc. Intuitively, mobile use
predictions of a user’s emotional state. The proposed model and its and background knowledge are associated with the emotions of
empirical development and validation are described in this paper. users. For instance, when they are happy, people may play mobile
phone games, or they may feel depressed if they live in a noisy en-
CCS CONCEPTS vironment. Several works were inspired by this, using Smartphone
sensing data for emotion detection.
• Human-centered computing; • Smartphones; • Non-
For instance, a multifactorial statistical model to understand
intrusive emotion detection;
everyday stress by comprehensively analyzing cell phone data and
weather conditions has been proposed in [4]. For cold-start emotion
KEYWORDS prediction using transfer learning, the authors in [5] used sensor
Mobile sensing, Mobile emotion sensing, Emotion Detection, Smart- data, APP usage information, and SMS content. However, their
phone Sensing approach is content-based involving privacy-sensitive user’s data
such as SMS content. In [6] the authors suggested a mood detection
ACM Reference Format:
system using the period of email, SMS, venue, and app use. Previous
Orestis Piskioulis, Katerina Tzafilkou, and Anastasios Economides. 2021.
studies have mainly investigated the identification of only one
Emotion Detection through Smartphone’s Accelerometer and Gyroscope
Sensors. In Proceedings of the 29th ACM Conference on User Modeling, Adap- emotion assuming that a person perceives only one emotional state
tation and Personalization (UMAP ’21), June 21–25, 2021, Utrecht, Netherlands. every moment. This study extends previous studies by relaxing this
ACM, New York, NY, USA, 8 pages. https://doi.org/10.1145/3450613.3456822 assumption and considers that a user may be in multiple co-existed
emotions called ‘compound emotions’. According to the theory of
Plutchik [7], a mixture of fundamental emotions would co-exist in
Permission to make digital or hard copies of all or part of this work for personal or a person at any moment. Compound facial feelings in human facial
classroom use is granted without fee provided that copies are not made or distributed
for profit or commercial advantage and that copies bear this notice and the full citation
expressions were also seen in the analysis of [8], [9]. For example,
on the first page. Copyrights for components of this work owned by others than ACM a compound emotional expression that incorporates basic feelings
must be honored. Abstracting with credit is permitted. To copy otherwise, or republish, of happiness and surprise is "happily surprised".
to post on servers or to redistribute to lists, requires prior specific permission and/or a
fee. Request permissions from permissions@acm.org. Furthermore, although there are several research studies on emo-
UMAP ’21, June 21–25, 2021, Utrecht, Netherlands tion detection through mobile devices, many of them apply obtru-
© 2021 Association for Computing Machinery. sive techniques of data collection (e.g., through external devices
ACM ISBN 978-1-4503-8366-0/21/06. . . $15.00
https://doi.org/10.1145/3450613.3456822
or wearables), eliminating their easy and practical implementation

Corrected Version of Record. V.1.1. Published July 20, 2021.


UMAP ’21, June 21–25, 2021, Utrecht, Netherlands Orestis Piskioulis et al.

in everyday user-Smartphone interactions [1], [10]. The literature 2.2 Emotion Recognition in Smartphones
review has shown that the accelerometer is the most frequently Mobile Sensing has been attracting research in recent years. The
used modality in non-intrusive mobile sensing methodologies. The productivity plateau is anticipated to be reached in around two or
front camera is frequently used as well; however, face tracking in five years, according to Gartner’s hype period of ‘’Emotion Detec-
mobile devices has raised several functional limitations (based on tion/Recognition” [22]. Nowadays, most people use Smartphones
the difficulty to efficiently capture the user’s full face) that might and researchers have shown that mobile sensing data from Smart-
affect the user experience during the data collection process [11]. phones can be adopted to infer and detect human mental well-being
Although researchers have implemented several prediction models such as stress, anxiety, depression, emotion and mood [9], [23]–
to identify human emotions through mobile sensor data, only a [26]. Several recent studies have managed to predict user emotions
few models have used solely non-intrusive methodologies and have through Smartphones in a non-obtrusive manner. Many of them
managed to achieve high accuracy scores [12]. Overall, the limited used accelerometer data in combination to other modalities like
number of studies, and their limitations imply that further research touch screen [27]–[29] or microphone, [30], [31]. However, their
is needed. collection of data lasted for several days or weeks and the feedback
Motivated by the increased need for further investigation of modalities included several more secondary ones like pedometer,
the prediction of emotions during the user interaction with mo- network, GPS, light sensor, keyboard etc.
bile devices, this study aims to investigate the recognition of two In a relevant study [12] implemented a multimodal methodology
basic human emotions (one positive and one negative) using ac- using accelerometer and keyboard data through a soft sensor to
celerometer and gyroscope sensors. Overall, our objective is to detect the emotions of happiness, anger, and neutral state. Their
apply a multimodal and non-intrusive methodology to predict user model implementation was based on Decision Trees, achieving
enjoyment and frustration during the use of a prototype gaming almost 93% accuracy.
app. In [29] the authors used a combination of accelerometer, gyro-
scope, and touch-screen data to predict two generic affective states
(positive and neutral). Their implemented support vector machine
2 THEORETICAL BACKGROUND model achieved a high accuracy score of about 87%. Most of the
2.1 Models for emotion and mood works that used an accelerometer unimodal approach were limited
in the detection of stress [32] or pleasantness [33] and generally
measurement
achieved lower scores of accuracies compared to the multimodal
There has been extensive research with regards to emotion and approaches.
mood in psychology, sociology, and neuroscience [13]. In general, As shown in [23], DeepMood detected bipolar affective disorder
“emotion” refers to the present instantaneous feeling, and “mood” on a multi-view neural network-based Smartphone using typing
refers to the longer-term normal feelings [14]. Such models can be dynamics and accelerometer sensor data. In [34] the author aimed to
used to quantify both instantaneous and long-term sensations, and assist mobile individuals when they need emotional help to connect
they can be used to quantify emotion and mood without limiting offline with friends. A multifactorial statistical model was proposed
differentiation. The Positive and Negative Affect Schedule (PANAS) by [4] to understand everyday stress by comprehensively analyzing
model is a widely used model to measure mood and broad affective cell phone data and weather conditions.
states [15]. Participants completing the PANAS are asked to rate Using mobile mobility trace data, the authors in [35] tracked
the degree to which each of the 20 emotions they encountered on a user depression states. The identification of positive and negative
5-point Likert scale range from “very slightly” to “very strongly”. emotions using wild mobile product sensors was suggested by
Participants can be asked how they feel right now or over longer [29]. Also, sensor data, APP usage information and SMS content
periods of time (e.g., over the past month), depending on the intent were used by the iSelf system [5] for emotion prediction using
of the measured emotion or mood state. transfer learning. However, these previous studies consider only
One popular model is Ekman’s six basic categories of emotions single emotion detection ignoring the fact that multiple emotions
[16]: happiness, sadness, anger, surprise, fear, and disgust. The Ek- can co-exist in time [36]. In our paper, we are trying to predict
man’s model is intuitive and understandable by the average users, two basic emotions (enjoyment and frustration), by making use of
and it allows multiple emotions to coexist with various intensity lev- Smartphones data from a collection of sensing information [37].
els. The Ekman’s emotion model has been widely adopted by many Overall, we notice there are not any methodologies based only
studies due to its high applicability [9], [17]. In different approach, on non-intrusive sensors of accelerometer and gyroscope. However,
Lövheim [18] compiled a list of 15 emotions that derive from the motivated by the above-mentioned studies, it seems worthy to
nine fundamental emotions of interest, joy, surprise, sorrow, fear, explore the success of accelerometer and gyroscope as the only
guilt, disdain, rage, and disgust. inputs to predict both positive and negative emotions.
Several emotions and affective states have been pointed out as
‘determinants’ in gaming or game-based learning experiences, since
they tend to affect a user’s overall experience and performance [19]. 2.3 Sensors
Overall, the emotions of enjoyment, frustration, confusion and anx- Modern Smartphones have a high number of sensors available, in
iety are outlined among the most frequently experienced emotions addition to software-based recordings of how the phone is used. For
during learning with a range of technologies, such as intelligent Android phones, the following sensors might be relevant for the
tutoring systems, simulations and serious games [20][21]. task of identifying emotions. In the present research and using the
Emotion Detection through Smartphone’s Accelerometer and Gyroscope Sensors UMAP ’21, June 21–25, 2021, Utrecht, Netherlands

Figure 1: Research Framework

appropriate technology we chose to record data from the following 3.1 Prototype App and Sensor Recording
sensors: We have developed a gaming app to serve as the prototype app
that participants will interact with during the experimental study.
⃝ Accelerometer (with gravity, motion and position detection); The prototype app has been developed in a way to be like business
⃝ Accelerometer Linear (without gravity, motion and position games that big firms are now using to provide an overview of their
detection); employees’ psychometric profile.
⃝ Gyroscope (affected by other vibrations, rotation detection). Overall, the designed game is a prototype framework that aims
to explore its users’ basic psychometric features. For this reason,
Both the accelerometer and the gyroscope provide measurements this app is based on similar commercial games (i.e., Skyrise City)
for each axis in a 3-dimensional coordinate system, and for this that are used massively in order to build psychometric profiles of
project we will focus on accelerometer readings. The output is a company’s potential employees.
vector consisting of the acceleration, in each of three directions x, For the creation and successful development of the prototype
y, and z. app we looked for a platform that could enable us to develop a fully
Ideally, the accelerometer of a phone lying on a flat surface compatible game with the recording of the sensors data. We finally
outputs the values A = {0, 0, ±g}. However, the accelerometer data decided to use the open-source game development platform, the
is not fully reliable. Even in an immovable phone, its accelerometer MIT App Inventor.
data does not remain constant, but they shift from one reading Using MIT App Inventor, we developed a game in the standards
to another. For this reason, accelerometers are usually not used of Skyrise City, capable of giving us a simple and short psycho-
for monitoring movement and orientation to detect emotions or metric game that will arouse the interest of users. Those kinds
affective data in general. Although affective data cannot be accessed of game-based assessments are grounded in validated personality
directly through the accelerometer, the forces exerted on the system and cognitive neuroscience frameworks and use real behavior data
are clearly generated. This enables us to determine the device’s to create a user’s psychometric profile. Through the challenges
orientation and movement [38]. faced by the user, a user profile can be created along six psycho-
logical characteristics: i) Aptitude, ii) Personal Style, iii) Cognition,
iv) Drive, v) Interpersonal Style, and vi) Thinking Style. Figure 2
3 MATERIAL AND METHODS depicts some interface snippets from the gaming app.
Our research methodology followed a multi-step approach. First, To record the accelerometer and gyroscope data, a sensor record-
a prototype psychometric gaming app was developed, and a free ing app was used. The MIT App Inventor and the Sensor Record
sensor log app was selected to be installed on the users’ Smartphone app ware both installed in a mobile phone. The app we used to
devices. Second, a self-reported measurement was designed accord- record data is a free Android application, Sensor Record, and gives
ing to the literature review to measure the examined emotions. the possibility to record 11 mobile sensors. You can easily monitor
Then, an experimental study was conducted on a set of volunteers your Smartphone’s sensor data with this app and save it on your
who were provided with all the necessary software (i.e., the pro- SD card or your computer for evaluation in csv-format.
totype gaming app installation, the sensor log app installation for
Android OS and the questionnaire link). After the data collection, 3.2 Survey Instrument
the sensor recorded data was processed and the selected features We developed a short questionnaire to collect users’ demographic
were extracted. Finally, a set of prediction models was developed, data, and previous gaming experience counting 10 items (four items
and their accuracy scores were compared. Figure 1 summarizes about demographics and six about the users’ familiarity with games
the research methodology to collect feedback from a sample of and technology). A second questionnaire (counting ten items in
participants who interact with the prototype Android app, and then total) was designed to measure the user’s emotions (on a five-point
process the data, extract the features to be fed into the model and Likert scale) after playing the game. Five items were used to mea-
finally train and evaluate the emotion recognition (ER) model. sure enjoyment and five items were used to measure frustration.
UMAP ’21, June 21–25, 2021, Utrecht, Netherlands Orestis Piskioulis et al.

parts of the survey. Two users discarded from the concentration


question and were excluded from the final sample. At this point,
we must point out that we have obtained data from 34 users from
the Huawei device while only 6 users have used their cell phones
for the game and the sensor recording system.
Regarding the users’ age, there were no significant differences,
since 30%, were 25-34 and 35-44 years old, 20% of the users were in
the age group 45-54, 17.5% in the group of 18-24 and finally 2.5%
of the participants were 55-64 years old. In terms of educational
level, 52.5% of the participants hold a postgraduate degree while
27.5% of the participants gained a bachelor’s degree. The remaining
percentage corresponds to a 15% for holders of a doctoral degree
and 5% for high school graduates. Regarding the time that the
participants spend in gaming activities, 42.5% of them spend from 4
to 10 hours a week, while 7.5 stated that spend more than 10 hours
a week. Also, 30% stated that they spend 1 to 3 hours, per week or
not at all. Finally, 7.5% stated that they spend more than 10 hours
in this activity. Finally, 27.5% of the participants stated that they
prefer sports games, 12.5% strategy games and 7.5% of users spend
Figure 2: Prototype gaming app- interface snippets
time on action, adventure, war, and car games.
In the beginning of the experimental process, the participants
The items were selected based on relative literature [39], and their were asked to manually install the required apps (game and sen-
wordings were carefully examined by two reviewers who are ex- sor log) in their Smartphones, or to use a given Smartphone (see
perts in the field of Technology Enhanced Learning, to eliminate the device characteristics in Table 2) that was already hosting the
any typos or inconsistencies. Among the questions, there was also required apps. Then, the users were asked to consent for their par-
a concentration question to filter and ultimately exclude those who ticipation and answer the first part of the online questionnaire and
answered by chance and gave contradictory answers. After col- play the prototype game. During playing the game, the "Sensor
lecting the data, the instrument was validated in terms of internal Record" application was running to monitor accelerometer and
consistency, according to the criteria of [40]. gyroscope sensing data. In the end, the participants answered the
As depicted in Table 1, the Cronbach’s Alpha value based on second part of the questionnaire to provide a self-reported feedback
standardized items for enjoyment, equals to 0.755 which is higher on their perceived enjoyment and frustration during playing the
than 0.7. Therefore, the data of our research about enjoyment are game. Table 2 shows the characteristics of the mobile phone that
considered reliable in terms of internal consistency. Similarly, the was used for the survey.
variable of frustration received a value of 0.709 or 70.9%, which is
considered reliable since it exceeds the reliability base of 0.7. Finally, 3.4 Data Overview
we can conclude that none of our questionnaire items predispose
A total of 1.367.049 sensor-based records were collected and pro-
the user to select or direct in one direction of answers and therefore
cessed for the classification of emotions, as depicted in Table 3.
the questions are considered inviolable and impartial, not guiding
Regarding the self-reported emotional data, the participants had
the user to certain pairs of answers.
to choose one of five levels for each dimension of emotions. While
such a data resolution might be interesting, to increase the classifi-
3.3 Participants and Procedure cation accuracy, the two highest levels and the two lowest levels
The experimental study was conducted on a set of users who volun- were combined resulting to three different classes on each emo-
tarily offered to participate. The volunteers were provided with all tional dimension for enjoyment and frustration: low, medium, and
the necessary software (i.e., the prototype gaming app installation, high as depicted in Table 4
the sensor log app installation for Android OS and the question-
naire link). Out of 75 users that were invited to participate (through
email –sent and direct invitation), a total of 42 users accepted to 3.5 Data (pre)Processing and Feature Selection
participate and finally 40 users (21 males, 17 females and 2 people The sensor data presents some difficulties when it comes to data
who preferred not to state their gender) successfully completed all quality:

Table 1: Cronbach’s Alpha (CA) values for the measured emotions

Reliability Statistics: CA Cronbach’s Alpha N of items


Enjoyment .755 5
Frustration .704 5
Emotion Detection through Smartphone’s Accelerometer and Gyroscope Sensors UMAP ’21, June 21–25, 2021, Utrecht, Netherlands

Table 2: Mobile Phone Characteristics

Mobile Phone Huawei P10


Android 8.0.0
RAM 3G
ROM 64GB
CPU 4x2.1GHz ARM Cortex-A53, 4x1.7GHz ARM Cortex-A53, Cores 8
GPU ARM Mali-T830 MP2, 900MHz, Cores 2
Dimensions 72x146.5x7.2 mm

Table 3: Amount of collected sensor data.

Sensor Amount of data (expressed in excel rows)


Accelerometer 450,886
Accelerometer Linear 455,549
Gyroscope 460,614
Total 1,367,049

Table 4: User-reported data & emotional classes

User-reported data [1, 5] Class


1-2 Low
3 Medium
4-5 High

⃝ The accelerometer (linear or not) and gyroscope are not For the success rate of a classifier, the choice of good features is
perfectly accurate, adding a little noise. essential. In this study, the selected features were extracted based on
⃝ When a user moves and plays with the phone, this causes suggestions of previous research studies [33]. The selected features
confusion to the sensors and it is difficult to compare the and their calculated values are depicted in Table 5. Three different
readings from individual axes. axes were measured by each sensor. Since in this work, we do not
⃝ If the mobile is not stable, this could make it more possible intend to consider the position of the telephone, it is reasonable
to jerk. to consider the total movement instead of the movement on the
⃝ The user will influence the recordings with gait features. individual axes.
⃝ It cannot determine whether the user moved or not during
the entire recording period. 4 PREDICTION MODELS
Firstly, we have cleaned out data by isolating: 4.1 Measured Emotions
⃝ Standard Types which the pandas recognized as out of the In this step we will try to predict two basic human feelings (Enjoy-
box. ment, Frustration) based on the mobile sensor data from accelerom-
⃝ Non-standard Types, that pandas have not automatically eter, accelerometer linear and gyroscope that we have tracked in
recognize as missing values. this research. To this effort we have constructed the following de-
⃝ Unexpected Types when a column of names contains unex- tecting emotion recognition models for the two basic emotions that
pected data types. we have targeted.
⃝ Summarizing Before doing that, in this research we will use the term emotion
⃝ Replacing, when making simple replacements mean value. We will define this term as the average value we re-
To accurately identify the accelerometer and gyroscope features ceived from the users of our research regarding the emotional state
that were essential to be included in the prediction model, we first that they perceived during playing the game. Next, we examine to
had to make following definitions. Given that we had to export what extent it has occur changes in the two basic emotional states
features from a huge amount of data, we set up a specific computing that we try to predict.
database. This was introduced as the calculation of the average of To give a simple example, when we receive a mean value of 3.8
every twenty observations so that we can proceed to the following for enjoyment, we concluded that our game had a positive impact
calculations. Practically, these 20 observations are made within a on this user because it makes them feel joy/entertainment. In simple
range of 200 milliseconds. In this way, we have been able to extract terms, their emotional state about this variable has changed from
accurate results for a variety of features. neutral, i.e., level of 3 (which we assume all users start in our
UMAP ’21, June 21–25, 2021, Utrecht, Netherlands Orestis Piskioulis et al.

Table 5: Extracted features from accelerometer and gyroscope sensors

Name Enjoyment Frustration


Mean Acceleration 3.77 2.61
Std Acceleration 0.034 0.031
Mean Peak Acceleration 0.12 0.23
Std peak Acceleration 0.0009 0.0012
Mean Jerk 0.002 0.005
Std Jerk 0.0001 0.0003
Mean Peak Jerk 0.07 0.13
Std peak Jerk 0.0004 0.0012
Mean step duration 0.01 0.024
Std step duration 0.0002 0.0005
Skewness -1.18 0.32
Kurtosis 1.46 2.00
Mean PSD 0.02 0.04
Std PSD 0.001 0.003

Figure 3: Self-reported emotion mean values, measured from 1 to 5 (N=40)

research purposes) to 3.8. Finally, this whole process has caused for both emotions of enjoyment and frustration. In fact, with this
a positive change to the term of enjoyment which is equal to 0.8. method we exceeded the average percentage of accuracy for the
Conversely, when a user’s frustration reported mean value 2.6, we detection of emotions using the sensors of the mobile and which
assume that there is a decrease from 3 to 2.6. This value reflects is 77% for this method [41]. Also, we must mention that the use
that this user did not feel any frustration after the whole process. of decision trees has been mainly used by researchers to predict
Figure 3 below depicts the self-reported mean values of perceived negative emotions like stress and positive emotions associated with
enjoyment and frustration for the sample of 40 participants. As enjoyment and happiness.
depicted, frustration meets lower values (of low or medium class We realize that the decision tree reached a high accuracy score
according to the labels in Table 4) compared with enjoyment (high of about 90% and is the most accurate model we have set up. It is
class according to the labels in Table 4), revealing an overall positive important to note that previous models found in the literature were
emotional experience of the participants. based on an accelerometer unimodal methodology, a multimodal
4.2 Performance of Prediction Models approach including several modalities or an approach including
The model development was conducted via the KNIME 4.2 Ana- camera feedback. In this research we have used only the accelerom-
lytics Platform. We developed a list of machine learning (ML) algo- eter and gyroscope and maybe this is one reason why some fore-
rithms as depicted in Table 6. We used 30% of the sample dataset to casting models did not provide us with the expected accuracy that
train the model and 70% of the dataset to test their prediction perfor- was met in previous emotion detection models [41].
mance scores. We have selected 70/30 as a training model based on
other researchers that have also used a high volume of data to train
emotion recognition models. Table 6 depicts the obtained results 4.2 Emotion Recognition Model
regarding the prediction success for the two targeted emotional
variables. More specifically, it was observed that by running the From the above analysis and presentation of the evaluation percent-
Decision Tree at KNIME, we obtained the optimal prediction rates ages of the models, we have selected the Decision Tree Algorithm
for the prediction of both basic emotions. We have chosen this
Emotion Detection through Smartphone’s Accelerometer and Gyroscope Sensors UMAP ’21, June 21–25, 2021, Utrecht, Netherlands

Table 6: Prediction models and performance scores

ML Algorithm Enjoyment Frustration


Linear Regression 65.13% 57.88%
Decision Trees 87.90% 89.45%
Logistic Regression 61.20% 42.12%
Support Vector Machine 72.08% 74.22%
Neural Network 77.16% 79.10%

method because it gives us the most accurate and continuous train- and infer complex emotional statements considering multiple basic
ing of the model that we have managed to achieve satisfactory emotions at one time.
accuracy in the data of this research. Moreover, the performance of the model depends heavily on
This model needs two types of data to work: training data of individual users. However, because of the wild
→X, Y, Z axis data nature of the study, a balanced dataset may not be possible. While
→Self-reported emotion values some filters and some techniques have been used to overcome the
It was observed that the more data we fed the model with, the imbalance problem, this should be carefully considered to obtain
more accuracy it returned. Secondly, this model uses the Decision balanced data during such application deployment. Machine learn-
Tree Predictor to try to predict the emotion value with the high- ing algorithms can be used specifically for unbalanced datasets [43],
est possibly accuracy. This value is based on the change in the [44].
mental state of the user in relation to the variables of enjoyment It is crucial to say that the deployment of emotion-sensitive ap-
and frustration. We conclude that this model can predict the self- plications on personal mobile devices challenges privacy issues. To
reported emotion value for enjoyment at 87.90% accuracy and the train prediction models, applications must record the sensor-data
self-reported emotion value for frustration at 89.45% accuracy. and collect information from their users. Because of their limited
This model was created with the ability to be continuously fed computational capability or battery power, mobile devices can dis-
back and try to increase its accuracy on a regular basis if it is charge the inferring task into cloud or another device. This means
equipped with the suggested sensor-extracted features. Thus, one that user data may be revealed. Location, device use, data from the
realizes that it can be used in similar research works that need activity, calls and communications contain all the information to
to estimate human emotion to provide with appropriate feedback. reveal the identity of a user. For this reason, we have anonymized
Finally, one realizes that mobile sensing emotion recognition models the personal data of our users.
need a lot of data to be successful, something that we need to Our model contributes significantly towards the non-intrusive
consider when dealing with the emotional change of users. detection of user emotions through Smartphone sensors. The sug-
gested model can be used in several ways. For instance, once we can
5 DISCUSSION AND CONCLUSIONS detect the emotional state of the, we can change the look and feel at
the interface in our effort to generate to the emotional state. More-
Overall, this study managed to apply a multimodal and non- over, emotion-based personalized services or empathetic agents can
intrusive methodology to predict user enjoyment and frustration be provided. As a fact, empathetic agents have been shown to be
during the use of a prototype gaming app. The results revealed effective in persisting positive emotions and altering an emotional
that the decision tree algorithm performs on the highest predic- state of fear to a neutral one [45].
tion model accuracies for both measured emotions. Our prediction Overall, this study reveals the efficiency of analyzing Smart-
model accuracy is 87.90% for enjoyment and 89.45% for frustration. phone sensor data to recognize user emotions of enjoyment and
The model was trained and evaluated based on sensor and self- frustration in a non-intrusive way. Our outcomes can contribute
reported data gathered from 40 users. This is a satisfactory sample to breakthroughs in practical applications in several fields, includ-
size compared to previous studies, but a larger sample could be ing affective user modeling and design of intelligent interfaces for
chosen in the future, as the data size is shown to add value to the learning, advertising/marketing, industrial engineering, and more.
model.
One limitation of this work is that we assume that in a user
session the emotion of the user does not generally change. While REFERENCES
we observe that 80% of the sessions are less than 4 minutes in [1] P. Wang, L. Dong, Y. Xu, W. Liu, and N. Jing, “Clustering-Based Emotion Recog-
our dataset, this assumption is probably untrue, but this may not nition Micro-Service Cloud Framework for Mobile Computing,” IEEE Access, vol.
8, 2020, doi: 10.1109/ACCESS.2020.2979898.
always be true. The design of advanced experience sampling method [2] E. Diener, S. Oishi, and R. E. Lucas, “Personality, Culture, and Subjective Well-
schedules to trace the emotional changes in the session can mitigate being: Emotional and Cognitive Evaluations of Life,” Annual Review of Psychology,
vol. 54. 2003, doi: 10.1146/annurev.psych.54.101601.145056.
this problem [42]. [3] J. J. Gross and O. P. John, “Individual Differences in Two Emotion Regulation
In addition, people often do not have a pure kind of emotion. Processes: Implications for Affect, Relationships, and Well-Being,” J. Pers. Soc.
For instance, an individual can be both excited and nervous when Psychol., vol. 85, no. 2, 2003, doi: 10.1037/0022-3514.85.2.348.
[4] A. Bogomolov, B. Lepri, M. Ferron, F. Pianesi, and A. Pentland, “Daily stress
a new process is applied. For a more accurate and comprehensive recognition from mobile phone data, weather conditions and individual traits,”
emotional inference service it is therefore essential to describe 2014, doi: 10.1145/2647868.2654933.
UMAP ’21, June 21–25, 2021, Utrecht, Netherlands Orestis Piskioulis et al.

[5] B. Sun, Q. Ma, S. Zhang, K. Liu, and Y. Liu, “ISelf: Towards cold-start emotion [25] Y. Suhara, Y. Xu, and A. S. Pentland, “DeepMood: Forecasting depressed mood
labeling using transfer learning with smartphones,” ACM Trans. Sens. Networks, based on self-reported histories via recurrent neural networks,” 2017, doi:
vol. 13, no. 4, 2017, doi: 10.1145/3121049. 10.1145/3038912.3052676.
[6] S. Du, Y. Tao, and A. M. Martinez, “Compound facial expressions of emotion,” [26] T. Stütz et al., “Smartphone based stress prediction,” in Lecture Notes in Computer
Proc. Natl. Acad. Sci. U. S. A., vol. 111, no. 15, 2014, doi: 10.1073/pnas.1322355111. Science (including subseries Lecture Notes in Artificial Intelligence and Lecture
[7] R. PLUTCHIK, “A GENERAL PSYCHOEVOLUTIONARY THEORY OF EMOTION,” Notes in Bioinformatics), 2015, vol. 9146, doi: 10.1007/978-3-319-20267-9_20.
in Theories of Emotion, 1980. [27] M. C. and K. Wac, “7515181 @ ieeexplore.ieee.org.” pp. 51–65, 2018, [Online].
[8] X. Bao, S. Fan, A. Varshavsky, K. A. Li, and R. R. Choudhury, “Your reactions Available: https://ieeexplore.ieee.org/document/7515181.
suggest you liked the movie: Automatic content rating via reaction sensing,” 2013, [28] S. Ghosh, S. Sahu, N. Ganguly, B. Mitra, and P. De, “EmoKey: An Emotion-aware
doi: 10.1145/2493432.2493440. Smartphone Keyboard for Mental Health Monitoring,” 2019, doi: 10.1109/COM-
[9] Y. Zhou, H. Xue, and X. Geng, “Emotion distribution recognition from facial SNETS.2019.8711078.
expressions,” 2015, doi: 10.1145/2733373.2806328. [29] A. Mottelson and K. Hornbæk, “An affect detection technique using mobile
[10] A. Wahid and F. Rasheed, “Emotion Detection Using Unobtrusive methods: An commodity sensors in the wild,” 2016, doi: 10.1145/2971648.2971654.
Integrated Approach,” 2019, doi: 10.1109/ICICICT46008.2019.8993118. [30] M. L. Barron-Estrada, R. Zatarain-Cabada, and C. G. Aispuro-Gallegos, “Multi-
[11] M. Khamis, F. Alt, and A. Bulling, “The Past, Present, and Future of Gaze-enabled modal recognition of emotions with application to mobile learning,” 2018, doi:
Handheld Mobile Devices: Survey and lessons learned,” MobileHCI 2018 - Beyond 10.1109/ICALT.2018.00104.
Mob. Next 20 Years - 20th Int. Conf. Human-Computer Interact. with Mob. Devices [31] M. Gjoreski, H. Gjoreski, M. Lutrek, and M. Gams, “Automatic Detection
Serv. Conf. Proc., pp. 1–17, 2018, doi: 10.1145/3229434.3229452. of Perceived Stress in Campus Students Using Smartphones,” 2015, doi:
[12] I. Zualkernan, F. Aloul, S. Shapsough, A. Hesham, and Y. El-Khorzaty, “Emotion 10.1109/IE.2015.27.
recognition using mobile phones,” Comput. Electr. Eng., vol. 60, pp. 1–13, 2017, [32] Maxhuni, Lecture Notes of the Institute for Computer Sciences, Social Informatics
doi: 10.1016/j.compeleceng.2017.05.004. and Telecommunications Engineering. 2017.
[13] J. Posner, J. A. Russell, and B. S. Peterson, “The circumplex model of affect: [33] A. F. Olsen, “Detecting Human Emotions Using Smartphone Accelerometer Data,”
An integrative approach to affective neuroscience, cognitive development, and Univ. Oslo Dep. Informatics, 2016.
psychopathology,” Dev. Psychopathol., vol. 17, no. 3, pp. 715–734, 2005, doi: [34] K. C. Herdem, “Reactions: Twitter based mobile application for awareness of
10.1017/S0954579405050340. friends’ emotions,” 2012.
[14] X. Zhang, W. Li, X. Chen, and S. Lu, “MoodExplorer: Towards Compound Emotion [35] L. Canzian and M. Musolesi, “Trajectories of depression: Unobtrusive monitoring
Detection via Smartphone Sensing,” Proc. ACM Interactive, Mobile, Wearable of depressive states by means of smartphone mobility traces analysis,” 2015, doi:
Ubiquitous Technol., vol. 1, no. 4, 2018, doi: 10.1145/3161414. 10.1145/2750858.2805845.
[15] J. R. Crawford and J. D. Henry, “The Positive and Negative Affect Sched- [36] Y. Gao, N. Bianchi-Berthouze, and H. Meng, “What does touch tell us about
ule (PANAS): Construct validity, measurement properties and normative data emotions in touchscreen-based gameplay,” ACM Trans. Comput. Interact., vol. 19,
in a large non-clinical sample,” Br. J. Clin. Psychol., vol. 43, no. 3, 2004, doi: no. 4, 2012, doi: 10.1145/2395131.2395138.
10.1348/0144665031752934. [37] S. Shah, J. N. Teja, and S. Bhattacharya, “Towards affective touch interaction:
[16] P. Ekman et al., “Universals and Cultural Differences in the Judgments of Facial predicting mobile user emotion from finger strokes,” J. Interact. Sci., vol. 3, no. 1,
Expressions of Emotion,” J. Pers. Soc. Psychol., vol. 53, no. 4, pp. 712–717, 1987, 2015, doi: 10.1186/s40166-015-0013-z.
doi: 10.1037/0022-3514.53.4.712. [38] A. Dzedzickis, A. Kaklauskas, and V. Bucinskas, “Human emotion recognition:
[17] R. Reisenzein, M. Studtmann, and G. Horstmann, “Coherence between emotion Review of sensors and methods,” Sensors (Switzerland), vol. 20, no. 3, pp. 1–41,
and facial expression: Evidence from laboratory experiments,” Emotion Review, 2020, doi: 10.3390/s20030592.
vol. 5, no. 1. 2013, doi: 10.1177/1754073912457228. [39] T. Ding, Yan; Zhao, “Emotions, Engagement, and Self-Perceived Achievement
[18] H. Lövheim, “A new three-dimensional model for emotions and in a Small Private Online Course,” J. Comput. Assist. Learn., vol. v36, no. n4, pp.
monoamine neurotransmitters,” Med. Hypotheses, vol. 78, no. 2, 2012, doi: 449–457.
10.1016/j.mehy.2011.11.016. [40] D. Gefen, “E-commerce: The role of familiarity and trust,” Omega, vol. 28, no. 6,
[19] R. Pekrun, “The control-value theory of achievement emotions: Assumptions, pp. 725–737, Dec. 2000, doi: 10.1016/S0305-0483(00)00021-9.
corollaries, and implications for educational research and practice,” Educ. Psychol. [41] K. Tzafilkou, N. Protogeros, and A. Chouliara, “Experiential learning in web devel-
Rev., vol. 18, no. 4, pp. 315–341, 2006, doi: 10.1007/s10648-006-9029-9. opment courses: Examining students’ performance, perception and acceptance,”
[20] A. C. Graesser, S. K. D’Mello, and A. C. Strain, “Emotions in advanced learning Educ. Inf. Technol., vol. 25, no. 6, 2020, doi: 10.1007/s10639-020-10211-6.
technologies.,” in International handbook of emotions in education., New York, NY, [42] H. Vachon, W. Viechtbauer, A. Rintala, and I. Myin-Germeys, “Compliance and
US: Routledge/Taylor & Francis Group, 2014, pp. 473–493. retention with the experience sampling method over the continuum of severe
[21] S. D’Mello and A. Graesser, “Dynamics of affective states during com- mental disorders: Meta-analysis and recommendations,” J. Med. Internet Res., vol.
plex learning,” Learn. Instr., vol. 22, no. 2, pp. 145–157, Apr. 2012, doi: 21, no. 12, 2019, doi: 10.2196/14475.
10.1016/j.learninstruc.2011.10.001. [43] C. Chen, A. Liaw, and L. Breiman, “Using Random Forest to Learn Imbalanced
[22] W. Goertz, “Hype Cycle for Human-Machine Interface, 2018,” Gartner, 2018. . Data | Department of Statistics,” 2004.
[23] B. Cao et al., “DeepMood: Modeling mobile phone typing dynamics for mood [44] Y. Tang, Y. Q. Zhang, and N. V. Chawla, “SVMs modeling for highly imbalanced
detection,” Proc. ACM SIGKDD Int. Conf. Knowl. Discov. Data Min., vol. Part F1296, classification,” IEEE Trans. Syst. Man, Cybern. Part B Cybern., vol. 39, no. 1, 2009,
no. October, pp. 747–755, 2017, doi: 10.1145/3097983.3098086. doi: 10.1109/TSMCB.2008.2002909.
[24] S. Servia-Rodríguez, K. K. Rachuri, C. Mascolo, P. J. Rentfrow, N. Lathia, and G. [45] C. N. Moridis and A. A. Economides, “Affective learning: Empathetic agents with
M. Sandstrom, “Mobile Sensing at the Service of Mental Well-being,” 2017, doi: emotional facial and tone of voice expressions,” IEEE Trans. Affect. Comput., vol.
10.1145/3038912.3052618. 3, no. 3, pp. 260–272, 2012, doi: 10.1109/T-AFFC.2012.6.

You might also like