Preference learning based approach to assess depressive symptoms and depression intensity in online forums

Internet users are subjected to a wide variety of threats. While some of these threats originate from criminals, other concerning risks may even stem from the users themselves. A growing body of research examines how social media usage and behavior relate to mental health conditions such as stress, anxiety, and depression. Current technologies developed to identify users at risk of these issues are largely reactive---that is, specific types of risks can only be detected by tracking Internet users when symptoms are already explicit and occurring. This reactive approach is insufficient, as we now possess the technological capability to go further and detect such problems at their earliest onset, before they become critical. The early risk prediction on the Internet (eRisk) project was initiated to explore the effectiveness of text analysis and machine learning techniques in the early detection of psychological risks among Internet users. This paper introduces a preference learning-based approach to assess different depression symptoms and identify their roles in explaining and predicting depressive behavior and its intensity. The proposed approach relies on the Dominance-based Rough Set Approach (DRSA), a typical preference learning method that offers higher explainability and interpretability than most existing machine learning techniques. The eRisk Test Collection, which includes data based on participants' responses to Beck's Depression Inventory (BDI), was used to evaluate the proposed approach. The results of the analysis were used to design and validate three strategies for the early detection of depression.