“Why does the computer say I am depressed?” Towards a better understanding of concepts in automatic depression detection

Abstract

This thesis explores the interpretability of automatic depression detection. Specifically, it investigates whether humanly understandable concepts can be gained from a varied array of machine learning models. Using the clustering of Bag of Words (BoW) representations through Principal Component Analysis (PCA), Latent Semantic Analysis (LSA) and Bag of Concepts (BoC), words are grouped into concepts, which are then examined for their interpretability. Concept Relevance Propagation (CRP), a post-hoc analysis technique to uncover latent concepts in neural networks, is also applied to a Convolutional Neural Network (CNN) which is trained on word embeddings. Two data sets are used: The Extended Distress Analysis Interview Corpus (EDAIC), a collection of semi-structured interviews, and D-Vlog, a collection of personal video blogs. The results show that deep learning approaches have the F 1 performance across the board, furthermore, on the D-Vlog dataset, PCA and LSA have a small negative effect on performance compared to a baseline BoW model, while dimensionality reduction through BoC causes a more significant decrease in performance. It is also shown that none of the simpler approaches are able to model the complexity of the EDAIC dataset. Moreover, the dis- covered concepts are compared to the symptoms of major depressive disorder as outlined in the DSM-V. The concepts discovered through CRP seem to be most meaningful as well as having some correlation to some symptoms of depression, with the concepts in both datasets, “Emotional Struggle” and “Being under Strain”, having a close correlation to feelings of depression. Concepts discovered through PCA and LSA, while coherent, are not able to be linked to any specific symptoms of depression. The concepts discovered through BoC do not have enough coherence to be compared to the symptoms of depression in any meaningful way.