Poster Preview
College
Bellini College of Artificial Intelligence, Cybersecurity, and Computing
Mentor Information
Seungbae Kim
Description
Suicide risk detection in online mental health communities poses significant challenges, particularly within bipolar disorder support groups where user interactions are diverse and complex. This study presents a foundational investigation using graph neural networks (GNNs) to model Reddit-based communities through both homogeneous and heterogeneous static graphs. In our homogeneous graph, where all users and interactions are treated uniformly, we extract key centrality measures—degree, closeness, betweenness, eigenvector, and PageRank—to identify influential users and understand the network structure. To capture the richer semantics of the heterogeneous graph, we derive meta-path-based centrality measures that leverage predefined meta-paths to uncover deeper relational patterns among different types of nodes and edges. We generate node embeddings by combining structural features from the graph with textual representations extracted using Sentence-BERT to evaluate baseline GNN architectures for downstream node classification. Our findings reveal distinct structural signals in each graph type and offer insights into risk-related patterns within the community. This work establishes a basis for future extensions incorporating temporal and dynamic graph modeling and emphasizes responsible AI practices, including user privacy and ethical data usage, throughout the research process.
Unveiling Suicide Risk in Bipolar Disorder Communities: Static Graph Analysis and Comparative Modeling on Reddit Networks
Suicide risk detection in online mental health communities poses significant challenges, particularly within bipolar disorder support groups where user interactions are diverse and complex. This study presents a foundational investigation using graph neural networks (GNNs) to model Reddit-based communities through both homogeneous and heterogeneous static graphs. In our homogeneous graph, where all users and interactions are treated uniformly, we extract key centrality measures—degree, closeness, betweenness, eigenvector, and PageRank—to identify influential users and understand the network structure. To capture the richer semantics of the heterogeneous graph, we derive meta-path-based centrality measures that leverage predefined meta-paths to uncover deeper relational patterns among different types of nodes and edges. We generate node embeddings by combining structural features from the graph with textual representations extracted using Sentence-BERT to evaluate baseline GNN architectures for downstream node classification. Our findings reveal distinct structural signals in each graph type and offer insights into risk-related patterns within the community. This work establishes a basis for future extensions incorporating temporal and dynamic graph modeling and emphasizes responsible AI practices, including user privacy and ethical data usage, throughout the research process.
