Online Social Event Detection via Filtering Strategy Graph Neural Network

Abstract

Nowadays, as a strongly time-dependent data type, the ubiquity of social media messages enables the detection and analysis of real-time events. Through the clustering of online posts concerning their topics, existing methods can quickly identify the current trends on social media, which helps discover marketing opportunities, prevent potential crises, etc. However, due to the diversity of social network users, the performance of current approaches is significantly affected by the long tail of random topics, which should be regarded as outliers in a clustering problem. Besides, current models are weak in detecting events that last for multiple days, which is common in real-world scenarios. Therefore, we propose the FS-GNN, a graph neural network based on a filtering strategy, for incremental social event detection in data streams. Our method uses heterogeneous information networks (HINs) to construct a social message graph, and we propose a centrality-based scoring mechanism to grade and filter noisy data before clustering. In addition, a message complement window is introduced to connect the same topic mentioned across multiple days for better clustering accuracy. Extensive experimental results demonstrate the superiority of FS-GNN over multiple baselines in both offline and online scenarios.