Graph Neural Networks (GNNs) have emerged as a powerful framework for graph representation learning. However, they often struggle to capture long-range dependencies between distant nodes, leading to suboptimal performance in tasks such as node classification, particularly in heterophilic graphs. Challenges like oversmoothing, oversquashing, and underreaching intensify the problem, limiting GNN effectiveness in such settings.
In this paper, we introduce WISE-GNN, a novel framework designed to address these limitations. Our approach enhances any GNN model by incorporating Wise-embeddings, which capture attribute proximity and similarities among distant nodes, thereby improving the representation of nodes in both homophilic and heterophilic graphs. Additionally, we propose a topological module that can be smoothly integrated into any GNN model, further enriching node representations by incorporating the topological signatures of node neighborhoods. Comprehensive experiments across various GNN architectures show that WISE-GNN delivers significant improvements in node classification tasks, achieving mean accuracy gains of up to 14% and 23% on benchmark datasets in homophilic and heterophilic settings, respectively. Moreover, WISE-GNN enhances the performance of various GNN architectures, allowing even standard GNNs to outperform SOTA baselines on benchmark datasets.