This document describes a study that aimed to build a mood sensor using smartphone usage patterns. The researchers collected smartphone usage data and self-reported mood entries from participants over several weeks. They then analyzed the data to develop models that could infer a user's mood based on their smartphone interactions. The best-performing models were personalized models for each user, but generally-applicable models and hybrid models were also explored. The goal is to provide mood sensing capabilities without additional hardware by leveraging the sensors already in smartphones.
This document describes a study that aimed to build a mood sensor using smartphone usage patterns. The researchers collected smartphone usage data and self-reported mood entries from participants over several weeks. They then analyzed the data to develop models that could infer a user's mood based on their smartphone interactions. The best-performing models were personalized models for each user, but generally-applicable models and hybrid models were also explored. The goal is to provide mood sensing capabilities without additional hardware by leveraging the sensors already in smartphones.
This document describes a study that aimed to build a mood sensor using smartphone usage patterns. The researchers collected smartphone usage data and self-reported mood entries from participants over several weeks. They then analyzed the data to develop models that could infer a user's mood based on their smartphone interactions. The best-performing models were personalized models for each user, but generally-applicable models and hybrid models were also explored. The goal is to provide mood sensing capabilities without additional hardware by leveraging the sensors already in smartphones.
This document describes a study that aimed to build a mood sensor using smartphone usage patterns. The researchers collected smartphone usage data and self-reported mood entries from participants over several weeks. They then analyzed the data to develop models that could infer a user's mood based on their smartphone interactions. The best-performing models were personalized models for each user, but generally-applicable models and hybrid models were also explored. The goal is to provide mood sensing capabilities without additional hardware by leveraging the sensors already in smartphones.
MOODSCOPE CLAIMS Feasibility of inferring mood from smartphone usage in energy efficient and privacy preserving system. Effect of mood on smartphone usage is personal. Lightweight and easy to deploy realization of moodscope. API for developers. INFORMATION LOGGED SMS Email Phone call Application Usage Web browsing Location The Science of Mood Mood Emotion Less Intense Lasts Longer Reaction to cumulative events More internal
Intense Spontaneous Reaction caused by a specific event Visible to others MEASURING MOOD Circumplex Mood Model
Discrete Categories
PANAS CIRCUMPLEX MOOD MODEL RELATED WORK
Recognizing emotions from voice and video
Alternate physiological signals
Leveraging signals from smartphones USER STUDY & DATA COLLECTION
Pre-Study Focus Group
Field Study
Post-Study Focus Group PARTICIPANTS 0 5 10 15 20 25 30 College and graduate students S/W Engineers Web Editor Salesman Teacher Others Occupation Males 66% Females 34% Beijing 78% US 22% Location PRE STUDY FOCUS GROUP Does application usage changes with mood? Do you communicate with people differently? Would they like to share their mood? Would they like their phone to adapt to their mood?
FIELD STUDY Mood Journaling Application Captures the users self reported mood. Input taken 4 times a day at 3 hr. interval Smartphone Interaction Logger Captures users smartphone interaction. For creating feature table.
MOOD JOURNALING APP. POST STUDY FOCUS GROUP
Is application easy to use? Is 4 mood input/day too frequent? How quickly they respond to input reminder? Characteristics of Mood Entries Response Rate
Daily Mood Persistence
Mood Distribution
RESPONSE RATE
0 10 20 30 40 50 60 70 80 1 2 3 4 or more P E R C E N T A G E ( % )
NUMBER OF RESPONSES DAILY MOOD PERSISTENCE
Average standard deviation- 0.41 Mood is fairly constant and slow to change.
Max. standard deviation- 1.73 Large mood swings do occur some days. MOOD DISTRIBUTION DESIGN OF MOODSCOPE MODELS
Daily Averages as Output
Usage Records as Input
Internal Model
Feature Table Characteristics DAILY AVERAGES AS OUTPUT USAGE RECORD AS INPUT Social Interaction Records Phone call, SMS, E-mails Routine Activity Records Browser History, App Usage, Location History Application Categorization Categories like built-in, games, office, social. INTERNAL MODEL Using mood averages as labels and usage record as feature table, Multi linear regression and SFS were applied to develop: Personalized Mood Model All-User Mood Model Hybrid Mood Model PERSONALIZED MOOD MODEL Pleasure Activeness Min. MSE 0.002 0.001 Avg. MSE 0.075 0.085 Max. MSE 0.176 0.215 Std. Deviation 0.050 0.062 Multi Linear Regression on each user data.
Requires Training. ALL USER MOOD MODEL One-Size-Fits-All Mood model.
For Pleasure Dimension Personalized All User Min. MSE 0.002 0.069 Avg. MSE 0.075 0.296 Max. MSE 0.176 0.79 Std. Deviation 0.050 0.175 HYBRID MODEL
Combines personalized and all user model approach. Accuracy depends on prioritization given to personalized data relative to data from general population.
FEATURE TABLE CHARACTERISTICS
Discriminative Features Number of occurrences in reduced table Number of statistically significant features Number of most Discriminative Features
Positively and Negatively Correlated Features MOOD INFERENCE ENGINE
PERFORMANCE API APPLICATIONS
Video & Music recommender systems. Social Sharing. Mood logger and browser. ADVANTAGES No extra hardware sensor. Light weight and power efficient 3.5mWH/day ~ 20 mins. of standby time. Application independent. Easy deployment.
LIMITATIONS Homogenous, small scale user population contribution to data set. Frequency of mood input. Not every factor can be captured. Lack of privacy.
Questions we could not answer Why not ?
What happens when no Wifi is available?
Is training of data a continuous process?
Difference between other and cannot categorize app categories. 38