DLAE: Dynamic Long-distance Attention Embedding for enhanced link prediction in complex networks

Dynamic link prediction is a crucial task in graph machine learning, with widespread applications in recommendation systems, traffic forecasting and transaction analysis. Existing dynamic graph neural networks often excel at modeling local short-distance dependencies, yet they suffer from insufficient capability or prohibitive computation cost when capturing long-distance dependencies in large-scale dynamic graphs. To address these issues, we propose Dynamic Long-distance Attention Embedding (DLAE), a lightweight dynamic graph learning method built upon the incremental updating paradigm for dynamic graphs. DLAE employs a dual-branch parallel architecture that enables efficient incremental learning and accurate link prediction on large-scale dynamic graph datasets by introducing a novel similarity-guided Lightweight Long-distance Attention (LLA) mechanism with virtual remote edges. Moreover, we propose a dedicated evaluation metric named Predicted Link Length (PLL) to evaluate the model’s ability to capture long-distance dependency information. Extensive experiments on real-world social and transaction networks show that DLAE achieves competitive performance against both classic and recent state-of-the-art methods. Our work demonstrates the significance of jointly modeling local and long-distance information for efficient and accurate dynamic link prediction.