A Generalised Depression Detection Model using Social Media Data

This thesis proposes two models for depression detection on social media: a Stacked Embedding Recurrent Convolutional Neural Network (SERCNN) and eXtreme boosting with berTopic (XT). The proposed models achieve competitive performance and maintain interpretability, which is crucial for clinical practitioners. The thesis investigates the generalisability of XT and proposes an extension, XT-2, which provides knowledge about the model's dominant features. The thesis also presents findings by optimising the effectiveness of the XT-2 framework, showing that accurate predictions can be made with relatively small amounts of data, reducing computational costs. Overall, the thesis provides efficient and interpretable models for depression detection on social media.