Upcoming Mood Prediction Using Public Online Social Networks Data: Analysis over Cyber-Social-Physical Dimension

Chaima DHAHRI  Kazunori MATSUMOTO  Keiichiro HOASHI  

Publication
IEICE TRANSACTIONS on Information and Systems   Vol.E102-D   No.9   pp.1625-1634
Publication Date: 2019/09/01
Online ISSN: 1745-1361
DOI: 10.1587/transinf.2018OFP0006
Type of Manuscript: Special Section PAPER (Special Section on Log Data Usage Technology and Office Information Systems)
Category: Emotional Information Processing
Keyword: 
mood prediction,  pearson correlation,  OSNs analysis,  autonomous system,  

Full Text: PDF(800.4KB)>>
Buy this Article




Summary: 
Upcoming mood prediction plays an important role in different topics such as bipolar depression disorder in psychology and quality-of-life and recommendations on health-related quality of life research. The mood in this study is defined as the general emotional state of a user. In contrast to emotions which is more specific and varying within a day, the mood is described as having either a positive or negative valence[1]. We propose an autonomous system that predicts the upcoming user mood based on their online activities over cyber, social and physical spaces without using extra-devices and sensors. Recently, many researchers have relied on online social networks (OSNs) to detect user mood. However, all the existing works focused on inferring the current mood and only few works have focused on predicting the upcoming mood. For this reason, we define a new goal of predicting the upcoming mood. We, first, collected ground truth data during two months from 383 subjects. Then, we studied the correlation between extracted features and user's mood. Finally, we used these features to train two predictive systems: generalized and personalized. The results suggest a statistically significant correlation between tomorrow's mood and today's activities on OSNs, which can be used to develop a decent predictive system with an average accuracy of 70% and a recall of 75% for the correlated users. This performance was increased to an average accuracy of 79% and a recall of 80% for active users who have more than 30 days of history data. Moreover, we showed that, for non-active users, referring to a generalized system can be a solution to compensate the lack of data at the early stage of the system, but when enough data for each user is available, a personalized system is used to individually predict the upcoming mood.