Learning the Explainable Semantic Relations via Unified Graph Topic-Disentangled Neural Networks

Graph Neural Networks (GNNs) such as Convolutional (GCNs) can effectively learn node representations via aggregating neighbors based on the relation graph. However, despite a few exceptions, most of previous work in this line does not consider topical semantics underlying edges, making less effective and learned between nodes hard to explain. For instance, current GNNs make us usually don’t know what is reason for connection network nodes, specific research topics cited article concerns among friends social platforms. Some methods have begun explore extraction recent related literature, but existing studies generally face two bottlenecks, i.e., either being unable explain mined latent relations ensure their reasonableness independence, or demanding textual content edges which unavailable real-world datasets. Actually, these issues are both crucial practical use. In our work, we propose novel Topic-Disentangled Network (TDG) address above at same time, explores from perspective contents. We design an optimized graph topic module handle features construct independent explainable semantic subspaces, then reasonable that correspond subspaces assigned each neighborhood routing mechanism. Our proposed model be easily combined with tasks form end-to-end model, avoid risk deviation representation space task space. To evaluate efficiency sufficient node-related conducted three public datasets experimental section. The results show obvious superiority TDG compared state-of-the-art models.