Dynamic graph representation learning based on multi-scale attention

Dynamic graph representation learning aims to learn informative and predictive node representations from evolving graph structures and node attributes. Although existing methods achieve promising results for modeling structural evolution and temporal dependencies, they often emphasize only a subset of the relevant structural or temporal information and lack a unified framework to jointly extract local topology, high-order structural similarity, node attribute dynamics, and multi-scale temporal dependencies. To fill this gap, we propose a Multi-scale Attention-based Dynamic Graph Network (MADGN) integrating structure-aware modeling and hierarchical attention. First, MADGN builds a time-aware hypergraph with structural descriptors and a temporal decay function to capture high-order correlations between structurally similar non-adjacent nodes. Next, Local Structure Attention and Temporal Feature Attention separately model fine-grained neighborhood topology and time-varying node attributes. We fuse these complementary embeddings via Long-Term Attention Fusion to capture long-range spatiotemporal evolutionary patterns, and design a dual-objective loss to jointly preserve connective proximity and structural similarity in the embedding space. Experiments on four real-world dynamic graph datasets validate that MADGN outperforms state-of-the-art baselines on dynamic link prediction, dynamic new link prediction, and node classification tasks. Our model achieves up to 6.17% relative AUC improvement for dynamic new link prediction, as well as maximum 14.5% Accuracy (ACC) and 7.9% AUC gains for node classification, while retaining efficient training overhead.