Leveraging Multiple Characterizations of Social Media Users for Depression Detection Using Data Fusion

Depression is one of the principal mental disorders worldwide, yet very few people receive the appropriate care needed due to the difficulty involved in diagnosing it correctly. Social networks have opened the opportunity to detect those users who suffer from this disease through the analysis of their posts. In this work, we propose using three types of characterizations (demographic, emotion, and text vectorization) extracted from the users' text and a fusion method for the detection of depressive users in the social network Reddit. Considering the diversity of each of the extracted characterizations, we adopted a Gated Multimodal Unit (GMU) as a fusion method. We compare this method against traditional data fusion methods and other methods that have used the same dataset. We found the proposed method improves Fl-score for the depressive class by 4% when combining these three characterizations. Showing the usefulness of characterizing user content and behavior for detecting depression and highlighting the impact that data fusion methods can have in this very relevant task.