DS-TAGCN: A Dual-Stream Topology Attentive GCN for Node Classification in Dynamic Graphs

Abstract

With the rapid growth of information, large amounts of graph-structured data have been generated. As an important task in graph-structured data research, node classification, which aims to classify nodes into different categories, has attracted a lot of attention from researchers in recent years. Real-life graphs are often dynamic whose graph topology and node attributes are constantly evolving. However, most of the studies focus on static graphs which can not capture the evolution of dynamic graphs. Node classification in dynamic graphs mainly has the following two challenges. First, it is difficult to effectively integrate modeling spatial and temporal features. Second, the evolution of dynamic graphs is located not only in node attributes but also in the graph topology. It is hard to learn the evolution of both aspects in the meantime. Besides, existing methods focus only on topological relations connected by explicit edges, while ignoring implicit topological relations that act in non-edge form. Implicit topological relations can help aggregate neighborhood features and further refine the modeling of node evolution patterns. To address these challenges and problems, we propose DS-TAGCN, a dual-stream topology attentive GCN for dynamic graph node classification. DS- TAGCN learns spatial-temporal features simultaneously by using a combination of GCN and LSTM. A dual-stream framework is designed to focus on the evolution of node attributes and graph topology, respectively. To mine the implicit topology, we propose TAGCN instead of GCN to model the implicit topological relations. Additionally, we incorporate a hierarchical attention mechanism in the network to automatically model the importance of different dimensional features. Extensive experiments demonstrate the effectiveness of DS-TAGCN.