Multi-order hyperbolic graph convolution and aggregated attention for social event detection.

Abstract

Social event detection (SED) aims to identify real-world events from large-scale social media streams and has become essential for applications in public safety, marketing analytics, and crisis management. However, the heterogeneous, hierarchical, and dynamic nature of social data poses fundamental challenges for conventional models built in Euclidean space, that struggle to capture non-Euclidean relational dependencies and higher-order event structures. To address these limitations, this study proposes the Multi-Order Hyperbolic Graph Convolution and Aggregated Attention (MOHGCAA) framework, which performs multi-order graph convolution in hyperbolic space while jointly modeling curvature-aware attention to capture both local and global dependencies. Extensive experiments conducted under both supervised and unsupervised settings show that MOHGCAA consistently outperforms existing state-of-the-art baselines across multiple datasets. The results highlight the model's robustness, scalability, and effectiveness in representing hierarchical and heterogeneous structures, providing a foundation for social event detection in non-Euclidean domains.