Intra- and inter-semantic with multi-scale evolving patterns for dynamic graph learning

Dynamics and heterogeneity are two principal challenges in recent graph learning research and are promising to solve many real-world applications better. Existing work usually decomposes heterogeneous graphs as different semantics via meta-path and learns each space representation separately. For dynamic graphs, they mostly learn the evolving patterns based on a general assumption that dynamic graphs change smoothly along with timestamps. Although much progress has been achieved, they still suffer from at least two open problems. First, different semantics are not always independent of each other because more information is reflected in their mutual interactions. Ignoring this latent influence might cause insufficiency in downstream tasks. Second, the smoothness assumption for graph evolving relies on temporal granularity. Specifically, two adjacent snapshots may change dramatically when the time interval is very coarse. To this end, this paper wishes to push forward this research area by learning interactions between different semantics and revisiting short-term changes within local areas and long-term changes at the global level. In particular, to address the first problem, we propose a novel cross-view mechanism to capture the mutual influence of different views. To solve the second problem, we construct a temporal graph with both short-term and long-term information and then design a graph attention network to learn multi-scaled features at both local and global levels. Extensive experiments on various graph tasks demonstrate the superiority of our model over state-of-the-art methods. (c) 2022 Elsevier B.V. All rights reserved.

Automated summary

Dynamics and heterogeneity are two major challenges in recent graph learning research, and addressing these challenges is crucial for real-world applications. Existing approaches usually decompose heterogeneous graphs into different semantics and learn separate representations for each space. However, there are still two open problems: the neglect of the mutual influence between different semantics and the reliance on smoothness assumption for graph evolving. This paper proposes a cross-view mechanism to capture the mutual influence and a graph attention network to learn multi-scaled features at both local and global levels, addressing these problems and demonstrating superior performance in various graph tasks.