Detecting adolescent depression in social media: A hierarchical ensemble learning approach

Abstract

As digital landscapes evolve, adolescents increasingly rely on social media platforms for self-expression, leading to the vast dissemination of their mental and emotional states. Amidst the physiological and psychological transitions characteristic of adolescence, there is a heightened risk of depressive disorders. These shifts, coupled with the dynamic nature of their online expressions, create a compelling case for detecting latent signs of depression within their digital footprints. This thesis delves into the intersection of Natural Language Processing (NLP) and adolescent depression detection, introducing an innovative hierarchical ensemble model tailored for the intricate task of identifying depressive markers in adolescent social media content. This model amalgamates traditional word embeddings with state-of-the-art pretrained models, encapsulating both sentence-level and word-level representations. Empirical validation, conducted on a unique dataset centered on adolescent depression detection, indicates the models superior efficacy over existing baselines. By offering an ensemble approach that captures the nuances of adolescent linguistic expressions, this research illuminates the potential for timely and non-intrusive interventions in adolescent mental health.