An Attention-Driven Graph Neural Network Approach for Mental Health Sentiment Detection

Abstract

Mental health issues such as depression, anxiety, stress, suicidal ideation, bipolar disorder, and personality disorders have become increasingly prevalent in the modern digital era. With the rapid growth of online platforms, individuals often express their emotional and psychological states through text-based content such as reviews, posts, and comments. Traditional artificial intelligence–based sentiment analysis systems primarily treat each text independently and fail to capture the complex semantic and relational dependencies that exist among mental health expressions. This limitation reduces the effectiveness of detecting high-risk mental health conditions, especially in overlapping and imbalanced categories. To address these challenges, this paper proposes a graph-based deep learning framework that integrates TF-IDF based text feature extraction with a Graph Attention Network (GAT) for mental health sentiment classification. TF-IDF is employed to transform textual data into numerical feature vectors representing the importance of words within the dataset, while a similarity-based graph structure is constructed using a K-Nearest Neighbor (KNN) approach to model relationships among text samples. The Graph Attention Network then applies attention mechanisms to effectively learn the relational dependencies between text instances and emphasize important mental health indicators. The proposed system classifies mental health data into seven categories: Normal, Depression, Suicidal, Anxiety, Stress, Bipolar Disorder, and Personality Disorder. Experimental results demonstrate that the proposed TF-IDF + GAT model achieves improved classification performance in identifying various mental health conditions and effectively captures relationships among textual data compared to traditional standalone machine learning approaches.