Higher-Order Graph Convolutional Embedding for Temporal Networks

Abstract

Temporal networks are networks that edges evolve over time. Network embedding is an important approach that aims at learning lowdimension latent representations of nodes while preserving the spatialtemporal features for temporal network analysis. In this paper, we propose a spatial-temporal higher-order graph convolutional network framework (ST-HN) for temporal network embedding. To capture spatialtemporal features, we develop a truncated hierarchical random walk sampling algorithm (THRW), which randomly samples the nodes from the current snapshot to the previous one. To capture hierarchical attributes, we improve upon the state-of-the-art approach, higher-order graph convolutional architectures, to be able to aggregate spatial features of different hops and temporal features of different timestamps with weight, which can learn mixed spatial-temporal feature representations of neighbors at various hops and snapshots and can well preserve the evolving behavior hierarchically. Extensive experiments on link prediction demonstrate the effectiveness of our model.