A Heterogeneous Propagation Graph Model for Rumor Detection Under the Relationship Among Multiple Propagation Subtrees

Pervasive rumors in social networks have significantly harmed society due to their seditious and misleading effects. Existing rumor detection studies only consider practical features from a propagation tree, but ignore the important differences and potential relationships of subtrees under the same propagation tree. To address this limitation, we propose a novel heterogeneous propagation graph model to capture the relevance among different propagation subtrees, named Multi-subtree Heterogeneous Propagation Graph Attention Network (MHGAT). Specifically, we implicitly fuse potential relationships among propagation subtrees using the following three methods: 1) We leverage the structural logic of a tree to construct different types of propagation subtrees in order to distinguish the differences among multiple propagation subtrees; 2) We construct a heterogeneous propagation graph based on such differences, and design edge weights of the graph according to the similarity of propagation subtrees; 3) We design a propagation subtree interaction scheme to enhance local and global information exchange, and finally, get the high-level representation of rumors. Extensive experimental results on three real-world datasets show that our model outperforms the most advanced method.