Modeling Heterogeneity with Flexible and Semantic Aligned Graph Neural Networks

Graph Neural Networks (GNNs) are particularly effective in machine learning tasks focused on graphs, notably in homophilous graphs where edges primarily connect nodes with identical labels. However, ensuring uniform outcomes in heterophilous graphs poses a challenge. Several studies have extended GNN architectures to address the unique characteristics of heterogeneous graphs. Nonetheless, these approaches often necessitate adaptability enhancement in managing feature combinations for both the central node and its neighboring nodes. To address this challenge, we introduce a Flexible and Semantic Aligned Graph Neural Network (FSAGCN). FSAGCN is based on two core principles. First, it independently encodes information from the central node and its neighbors, combining them to produce a fused output. Second, it aggregates nodes of the same class, aligning them with semantic information from both the feature and neighbor perspectives to impose constraints on the conversion of central nodes and neighbors during the propagation of the graph structure. We conducted comprehensive experiments with eight real-world datasets, demonstrating that FSAGCN outperforms state-of-the-art (SOTA) baseline models.