Invariant Subgraphs for Cross-Domain Fake News Detection via Causal Disentanglement

Abstract

The spread of misinformation through social media poses significant threats. Recent models using text and graph features have shown promising results in specific misinformation detection scenarios. However, these data-driven models heavily rely on training data that share similar distribution with inference data, limiting their applicability to misinformation from emerging or previously unseen domains, known as out-of-distribution (OOD) data. Tackling OOD misinformation is a challenging yet critical task. To address the challenge, we propose the Causal Subgraph-oriented Domain Adaptive misinformation Detection (CSDA) model. Based on a causal analysis, CSDA extracts invariant substructures from news propagation graphs that generalise to OOD data, using a graph neural network-based mask generation process. It uses refined training objectives to ensure high-quality subgraphs. It is further powered by contrastive learning for few-shot scenarios, where a limited amount of OOD data is available for training. Extensive experiments on public social media datasets demonstrate that CSDA effectively handles OOD misinformation detection, achieving a 1.23%∼12.23% accuracy improvement over other state-of-the-art models, covering OOD news domains in politics, entertainments, health, etc.