Suicide ideation detection via social and temporal user representations using hyperbolic learning

Social media data exhibits complex hierarchical structures and temporal dependencies that are challenging to capture with traditional Euclidean representations. We propose a novel approach for suicide ideation detection that leverages hyperbolic geometry to learn rich social and temporal user representations. Hyperbolic spaces are particularly well-suited for modeling hierarchical relationships and have shown promise in capturing complex network structures. Our method embeds users in hyperbolic space, where social connections and temporal patterns can be more naturally represented. The approach jointly models social relationships between users, temporal evolution of their mental health states, and textual content of their posts. By operating in hyperbolic space, our model can capture the tree-like and hierarchical nature of social interactions and temporal dependencies more effectively than Euclidean methods. Experimental results on real-world social media datasets demonstrate that hyperbolic representations lead to significant improvements in suicide ideation detection performance, particularly in capturing long-range temporal dependencies and complex social dynamics.