GraphDAFI: A graph representation learning framework with degree-aware feature interaction for node classification

Graph neural networks (GNNs) have been widely studied to handle graph-structured data due to their superior learning capability. Despite the successful applications of GNNs in many areas, their performance suffers heavily from the imbalanced degree distribution (long-tail issue). Most prior studies tackle this issue by graph augmentation, which explicitly increases the communication among nodes by optimizing original topology. In this paper, we employed the perspective of taylor interaction to explore the long-tail issue, and analyzed that there is insufficient interaction between low-degree nodes and their neighbors. In detail, we proposed a novel GNN framework named with degree-aware feature interaction (GraphDAFI), in order to bridge the gap of neighborhood aggregation between head-node embeddings and tail-node embeddings. GraphDAFI comprises two collaborative modules: adaptive feature interaction and degree-aware neighborhood transfer. Adaptive feature interaction leverages node embeddings of the current layer and interactions of the historical layer to perceive potential local information. Then, a unified feature encoder was designed that enhances the interaction to increase the model's generalization ability. To inject relevant information into low-degree nodes, a degree-aware neighborhood transfer was developed, which updates the node-edge adjacency matrix through a degree-aware strategy to achieve knowledge transfer. Experimental results demonstrate that GraphDAFI achieves excellent performance in semi-supervised node classification compared with the state-of-the-art models.