Burst-Aware Community Detection

Event-based temporal networks provide a natural representation for modeling interactions in digital systems. However, identifying persistent communities remains challenging due to bursty and irregular activity patterns. This paper proposes a burst-aware framework that integrates adaptive temporal segmentation, co-activity graph construction, representation learning, and structural impact analysis to detect persistent and influential communities. Unlike approaches based on fixed temporal windows, the proposed method aligns network construction with burst episodes, preserving meaningful coordination periods while reducing temporal noise. Community influence is estimated through a unified score combining structural centrality, temporal persistence (TSE), and robustness under node removal (Hide-1). The framework is evaluated on two datasets with distinct interaction regimes: a collaborative learning environment (OULAD) and a competitive programming environment (Online Judge). Results show that burst-aligned segmentation improves ranking stability, persistence, and recoverability when compared with fixed-window, aggregated, and random segmentation baselines. These findings highlight the importance of burst-aware temporal modeling for detecting persistent communities in event-driven systems.