Temporal Blocks with Memory Replay for Dynamic Graph Representation Learning

Le résumé fourni par la source

Dynamic graph representation learning (DGRL) aims to model the temporal evolution of graph structure and attributes, thereby generating low-dimensional node representations at different time steps. Most prevailing snapshot-based methods construct snapshots independently in time, assigning each interaction to a single snapshot. However, such a design limits the ability to capture long-range temporal patterns, leading to the forgetting of prior interactions and reducing the capacity of the model to recognize causal dependencies across events. To address this issue, we construct temporal blocks with the memory replay mechanism by sequentially merging several adjacent snapshots to capture long-range temporal patterns and causal dependencies over time. Building on this, we propose a novel dynamic graph representation learning model named TBD. Specifically, the model first encodes each temporal block using a graph neural network (GNN), and then captures cross-block dynamics through a Multi-Feature Gated Recurrent Unit (MF-GRU) that incorporates structural embeddings and a feature-aware gating mechanism to adapt to evolving graph structures. Furthermore, we introduce a Structure-Aware Node Smoothness Constraint (SA-NSC) to enforce temporal consistency while retaining adaptability to structural changes. Extensive experiments on multiple real-world datasets demonstrate that TBD consistently achieves superior performance, validating its effectiveness and robustness.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.