Not All Neighbors are Temporally Relevant: An Adaptive Neighborhood Aggregation Framework for Dynamic Graph Learning

Bingce Wang, Weiping Li, Tong Mo, Xiang Yuan, Xu Chu, Liwen Zhang

Abstract: Discrete-Time Dynamic Graphs (DTDGs) are represented as a sequence of static graphs, i.e., graph snapshots, in which nodes and edges may change from snapshot to snapshot. Most current Graph Neural Networks (GNNs) for DTDGs follow a paradigm where the sequence encoder and the graph encoder are combined within a fixed receptive field. Although this paradigm can capture the temporal evolution of subgraphs, it may also include temporally irrelevant neighborhoods because of the unchanged receptive field. The method of determining a temporally correlated neighborhood in DTDGs remains unexplored. To fill the gap, we innovatively propose GraphANA, an Adaptive Neighborhood Aggregation framework for dynamic graph embedding. For the first time, we transfer the prevailing concept of “one node one receptive field” to the DTDGs. Specifically, GraphANA first defines multi-hop neighborhood propagation within a shortest-path kernel function to derive neighborhood representations for each snapshot. Then, GraphANA employs a hop-aware sequence encoder to generate time-series tokens that capture the temporal properties of each neighborhood’s feature. Finally, a dual-stage adaptive neighborhood aggregation is conducted to eliminate irrelevant neighborhoods and to aggregate each time-series token from the remaining neighborhoods with temporal relevance scores. Extensive experiments demonstrate that GraphANA achieves competitive performance on downstream tasks.

External IDs:dblp:conf/dasfaa/WangLMYCZ26