Multiplex graph prompt collaboration for open-set social event detection

Abstract

Social event detection (SED) aims to detect event types from social media messages that reflect group behavior and public concerns. Owing to the dynamic evolution nature of social media, newly occurred messages may belong to unseen event types. Recently emerged open-set SED methods introduce graph neural networks (GNN) for feature encoding and iteratively retrain model parameters via the learned pseudo-labels to adapt to new, unknown events. However, retraining the entire model with these noisy pseudo-labels inevitably causes overfitting and risk erasing information learned from known events. In this paper, we propose a novel framework to tackle the open-set SED problem by leveraging the advantages of graph prompt learning (GPL) in its fast adaptation to new data with minimal parameter modifications, termed Multiplex Graph Prompt Collaboration (MGPC for short). Specifically, MGPC introduces three types of graph prompts to adapt a pre-trained GNN model from old to new message graphs quickly without extensive retraining. To address the distribution shifts issue in node content, graph structures, and temporal context as new data emerges, we introduce a graph adaptation prompt and a temporal prompt, which manage information flow from old to new graphs under evolving temporal context. We further introduce a prototypical-based supervised training loss with a lightweight event embedding prompt, which facilitates quick adaptation to new event class distributions while retaining previously learned information with minimal parameter changes. These adopted prompts are fine-tuned using pseudo-labels generated according to the entropy-based uncertainty scores concerning the known classes, supplemented by an unsupervised contrastive learning component to improve inter-class discrimination for unknown events. Extensive experiments on real-world benchmarks demonstrate the effectiveness of the proposed MGPC framework in comparison to existing SED methods.