Enhancing Decoder for Graph Automorphic Node Equivalence in Link Prediction

Graph Neural Networks (GNNs) have limitations in distinguishing between automorphic nodes, which are structurally identical within a graph, leading to difficulties in predicting links between these nodes. Existing methods typically address this issue by designing sophisticated GNN encoders, such as labeling nodes uniquely or limiting the analysis to specific subgraphs, which transforms the link prediction problem into a binary subgraph classification task.However, the role of the decoder, which transforms encoded node pairs into final link predictions, has been largely neglected. To bridge this gap, we propose DBLP (Decoder Boosting for Link Prediction), a simple yet universal framework that enhances the decoder to address the automorphic node problem. We demonstrate that even a basic graph encoder integrated into DBLP can achieve comparable or superior results to existing methods, highlighting the significant yet often overlooked impact of decoders.Our experiments demonstrate that integrating a simple graph encoder such as GCN with the DBLP framework can lead to performance improvements of approximately 17%, with no obvious computational overhead. Extensive evaluations across diverse datasets confirm both the effectiveness and the generalizability of DBLP, highlighting its potential in improving GNN-based link prediction.