Visualizing Language : A Chernoff Face Approach for Interpreting Embeddings From Online Mental Health Posts

This thesis introduces a unique approach for visualizing mental health conditions from social media data, focusing on a structured pipeline involving data preparation, feature extraction, Chernoff Face visualization, and clinical validation. Initially, the pipeline begins with dataset acquisition and cleaning, reducing the dataset from 15,744 to 2,621 posts, emphasizing the importance of a clean and relevant dataset for analysis. Prompt engineering further refines the data by transforming posts and questions into prompts indicative of specific mental health issues.

Feature extraction is a critical step where various techniques like attention-weighted features, feature averaging, max-pooled features, min-pooled features, and concatenated features are employed to distill the essence of textual data into meaningful patterns. The integration of a Sparse Autoencoder selects the optimal feature selection method based on the lowest reconstruction error, leading to a focused dimensionality reduction from 50265 to 256 features.

The visualization phase employs PCA (Principal Component Analysis), t-SNE (tDistributed Stochastic Neighbor Embedding), and UMAP (Uniform Manifold Approximation and Projection) for further dimensionality reduction, resulting in compact feature sets that are then transformed into Chernoff Faces, a novel method for visualizing mental health conditions through facial expressions. Clinical validation involves a preliminary evaluation by psychiatry volunteers, focusing on the interpretability and relevance of these visual representations across different mental health conditions. This initial feedback guides the refinement of Chernoff Faces, aiming for enhanced accuracy and clinical utility.

Subsequently, more detailed analysis and visualization of mental health conditions are carried out, showcasing the comprehensive capability of the developed pipeline in not just depicting but also accurately classifying various mental health conditions through visual means. The study reveals significant findings, including psychiatrists' preferences for certain visualization techniques and the potential clinical implications of integrating such innovative tools into psychiatric evaluation.

The conclusion emphasizes the importance of further evaluating the combined feature method, which incorporates PCA, sentiment analysis, and mental issue encoding for a refined visual representation. This future work aims to validate the approach with an extended study involving psychiatry volunteers, promising to enrich the psychiatric diagnostic toolkit with visually intuitive and clinically relevant methods.This thesis not only contributes to the field of computational analytics and psychiatric evaluation but also highlights the potential of advanced visualization techniques in enhancing the understanding and diagnosis of mental health conditions, fostering a more empathetic and insightful approach to mental health care.