Emotional Insights and Predictions Using Non-Intrusive Smartphone Activities

Abstract

Emotion identification is a complex research area that can enable unique multi-device experiences. Smartphones, the dominant mode of communication, can aid in emotion prediction. However, there is a lack of datasets with precise ground truth labels based on user smartphone behavior due to challenges associated with dataset annotation. Present annotation techniques rely either on self-reporting or recording on desktop applications, which is less natural. In this work, we address these issues by devising a user-centric approach to collect and annotate user data in a non-intrusive way on smartphones. We derive insights from the annotated data comprising behavior, emotion, and personality. The data consists of categorical features that do not include personally identifiable information, thus preserving user privacy. We validate the annotated data by an emotion prediction model using the Random Forest classifier, achieving an accuracy score of 67.73%. Further, we achieve an accuracy of 77.95 % on sentiment prediction (positive, negative, and neutral) using the Support Vector Machine (SVM) classifier.