DEMO: Heterogeneous Multilayer Density infused Entropy-Modularity Optimization for Unsupervised Social Event Detection

ACL ARR 2025 May Submission7491 Authors

Abstract: Event detection from social streams is an essential component of monitoring real-world incidents with applications in disaster monitoring, health surveillance and public opinion analysis, among others. Social media generates information streams containing heterogeneous attributes, such as names, places, and times, which often exhibit noise as the same entities may belong to different events, making detection challenging. The present paper introduces an unsupervised event detection model DEMO (Heterogeneous Multilayer \textbf{D}ensity infused \textbf{E}ntropy-\textbf{M}odularity \textbf{O}ptimization). DEMO judiciously optimize both entropy and modularity to deal with the noise arising from multiple heterogeneous interactions. This allows better classification of events from social streams. The method is aided by a community detection algorithm, mCOMM, which infuses heterogeneous multilayered density-based community participation information into the optimization pipeline. Extensive experiments support our model's superior performance, surpassing SOTA methodologies with a maximum gain of $110.7\%$ in ARI, $20.2\%$ in AMI and $19.7\%$ in NMI for publicly available datasets.

Paper Type: Long

Research Area: Machine Learning for NLP

Research Area Keywords: Density, Entropy, Modularity, Event Detection, Heterogeneous, Multilayered, Streaming

Contribution Types: Model analysis & interpretability, NLP engineering experiment, Theory

Languages Studied: English, French

Submission Number: 7491