Toward Effective Multi-Domain Rumor Detection in Social Networks Using Domain-Gated Mixture-of-Experts

A research team has published a new paper introducing PerFact, a novel approach to detecting rumors across multiple domains on social media platforms like X. The core problem they address is the performance degradation of single-domain detection models when faced with new topics, due to shifts in lexical patterns and information propagation dynamics. To solve this, they first created a large-scale, high-quality dataset of 8,034 annotated posts (with an annotator agreement κ=0.74), categorized into rumor and non-rumor classes. They then proposed a new model architecture designed specifically for this multi-domain challenge.

The proposed model employs a domain-gated Mixture-of-Experts (MoE) framework. Each 'expert' in the mixture combines a Convolutional Neural Network (CNN) to capture local syntactic features and a Bidirectional Long Short-Term Memory (BiLSTM) network to understand long-range contextual dependencies in the text. A domain gate dynamically aggregates the insights from these multiple experts based on the input. By leveraging both the textual content and publisher metadata, the model classifies posts with high accuracy. Evaluations show it achieves a state-of-the-art F1-score of 79.86% and an accuracy of 79.98% in multi-domain settings, outperforming previous specialized models and demonstrating robust generalization across topics like politics, health, and entertainment.