Multi-feature graphs and contrastive learning for rumor detection on social media

Social media platforms are rife with rumor texts, which, if not mitigated in time, can lead to widespread misinformation. Traditional rumor detection methods often rely on textual graphs that fail to capture polysemy and struggle with the informal, slang-rich nature of social media content. To address these challenges, we propose a novel multi-feature graph-based rumor detection model for social media. Specifically, we construct four complementary graph structures: a raw text co-occurrence (CO) graph, a part-of-speech (POS) graph, a named entity (NE) graph, and a semantic dependency (SD) graph. To effectively integrate these heterogeneous graphs, we design a gated multi-layer perceptron (MLP)-based fusion mechanism that adaptively aggregates salient features from each graph. Furthermore, to enhance the model’s discriminative capability, we introduce a cluster-driven contrastive learning strategy that encourages features of similar rumors to be closer in the representation space while pushing apart dissimilar ones, enabling automatic clustering of semantically similar rumors. Extensive experiments on three widely used rumor detection datasets demonstrate that our model achieves superior performance, even with only 1% of the data used for training, significantly outperforming existing baselines in terms of accuracy. • A multi-feature graph architecture captures diverse social media features. • A gated MLP module adaptively integrates signals from heterogeneous graphs. • Cluster-driven contrastive learning enhances representations via pseudo-labels. • Robust performance in low-resource settings using only 1% of training data. • SOTA results achieved on three public Chinese rumor detection datasets.