Text-Rich Graph Neural Networks With Subjective-Objective Semantic Modeling

Graph Neural Networks (GNNs), which obtain node embeddings by attribute propagates along graph topology, exhibit significant power in graph-structured data mining. However, graphs in the real world are usually text-rich, where the text can not only be represented as node attributes, but also contains valuable objective semantic structures. Moreover, the graph topology also exhibits complex subjective semantic structures, especially the heterophily where nodes from different classes are prone to build connections, making existing GNNs that work under the assumption of homophily incapable to realize generalization. To tackle aforementioned limitations, we design a new text-rich graph neural network from a unified perspective, namely SO-GNN. It can effectively enhance the expressive power of GNNs by modeling the implicit but informative subjective-objective semantics underlying the text-rich graphs. Specifically, we first introduce a new constrained Markov matrix with well-defined probabilistic diffusion dynamics to guide information propagation, where the neighbors are more appropriate and indicative in providing both local and global subjective semantics. We then construct a flexible heterogeneous text graph to gain a deeper insight into objective semantics, providing indispensable information for learning node embedding. Finally, we unite subjective and objective semantics in an end-to-end manner, so that the model can fully utilize the most relevant information for downstream tasks. Extensive experiments across various text-rich graphs with low-to-high homophily demonstrate the effectiveness and flexibility of the proposed SO-GNN over state-of-the-arts.