Computer Science · Artificial Intelligence
Heterogeneity Modeling with Flexibility-based Graph Neural Networks
Jisheng Qin, Tao Tao, Zijie Liu, Haibo Chen, Shenghui Zhao
See on the mapAbstract
Graph Neural Networks (GNNs) have demonstrated robust performance in various graph-based machine learning tasks, especially in homophilous graphs, where edges tend to connect nodes with the same label. However, achieving consistent performance on heterophilous graphs remains challenging. Several works have extended GNN architectures to address the distinct characteristics of heterogeneous graphs. However, these methods often need to enhance the flexibility of handling feature combinations for the central node and its neighbors. In this paper, we introduce a novel approach to modeling heterogeneity with Flexibility-based Graph Convolutional Networks (FGCN). FGCN consists of two primary components: first, it independently encodes the central node and its neighbors, fusing them to obtain an output representation. Second, it aggregates nodes with the same label from a feature and neighbor perspective, imposing constraints on the transformation during graph structure propagation. This approach enhances the flexibility of combining the central node and neighbor representations. We conducted extensive experiments on eight real-world datasets, demonstrating that FGCN outperforms state-of-the-art (SOTA) baseline models.