Inductive Representation Learning of Temporal Pattern Subgraphs in Temporal Networks

Dongdong Chen, Qiyao Huang

Keywords: graph representation learning; time series analysis

Abstract: Subgraph structure learning on graph neural networks (GNN) has attracted considerable attention recently because of its capacity to encode high-level graph structural features. Temporal network representation learning has been used in many real-world dynamic systems that usually evolve according to some temporal patterns, such as the triadic closure laws in social networks. Inductive representation learning of temporal networks should be able to capture these temporal patterns and further apply them to nodes which were not discovered during the training. Previous work neglected to extract these patterns, or could not be applied to downstream tasks because of the high time complexity of matching. In this paper, we design the strategy for capturing two types of prevalent temporal patterns. We propose the TPSN framework for inductive temproal pattern learning, in which we perform adjacent edge reconstruction on the extracted subgraphs, thereby improving the learning efficiency of temporal triadic closure laws and reducing the possibility of oversmoothing. Furthermore, we use multi-subgraph contrastive learning to achieve higher accuracy with fewer negative samples. Our proposed method outperforms baselines in all three downstream tasks and maintains acceptable time complexity. Ablation experiments also validate the effectiveness of our proposed model and module.

Primary Area: learning on time series and dynamical systems

Submission Number: 7198