CLRDMMF: A Contrastive Learning-based Rumor Detection Model with Multi-feature Fusion

Current social media platforms provide conditions for the widespread propagation of rumors, which can harm public opinion and even social stability. To automatically detect rumors, rumor detection methods employ graph neural networks to extract structural features in the rumor propagation process and achieve good classification performance. However, these methods have the following three limitations, i.e., the model architecture, information utilization and training methods. In order to overcome these limitations, in this paper, we design a Contrastive Learning-based Rumor Detection Model with Multi-feature Fusion (CLRDMMF). First, CLRDMMF designs a powerful GNN propagation structural feature extractor called Cross-directional Attention Graph Isomorphism Network (CA-GIN). Then, CLRDMMF extracts and fuses multiple kinds of features from the propagation tree, source text and Kernel Subtree. After that, we build the classification model with both the supervised and unsupervised contrastive learning. Finally, we conduct evaluation experiments and demonstrate that the proposed method outperforms the best baseline with 1.9% and 2.8% in accuracy on two datasets.