Probabilistic Label-Aware Aggregation-Based Heterophilous Graph Neural Network

In recent years, graph neural networks (GNNs) have demonstrated remarkable potential in graph representation learning. Most existing studies are based on the homophily assumption. However, real-world graphs are predominantly heterophilic, characterized by connections between nodes of different categories, where traditional GNNs tend to underperform. While several existing heterophilic GNNs attempt to address these challenges, the majority fail to fully leverage neighborhood label information of target nodes. In this work, we propose PLA-HGNN (Probability Label-Aware Heterophilic Graph Neural Network), a novel approach for node representation learning on heterophilic graphs. PLA-HGNN generates probabilistic node labels based on both feature matrices and adjacency matrices, then explicitly partitions a target node's neighbors according to these probabilistic labels. This enables our guided graph neural network to separately aggregate neighborhood information from different categories. Furthermore, to enhance generalizability, we introduce a learnable parameter to balance the contributions of different nodes during the aggregation process. We conduct extensive experiments on representative homophilic and heterophilic datasets, with results demonstrating PLA-HGNN's superiority in semi-supervised node classification tasks.