Network Representation Learning Method Based on Spatial-Temporal Graph in Dynamic Network

Network representation learning, which aims to learn the low-dimensional representations of vertices, has attracted considerable research efforts recently. Most existing network representation learning methods mainly focus on static networks, which extract and condense the network information without temporal information. However, in the real world, networks keep evolving and the evolution process contains a lot of important information. But it is difficult to quantify the evolution of node structure in dynamic networks. To address this problem, we propose a dynamic network representation learning method, call it as RWR-STNE. The proposed framework makes use of structural properties of networks at current and previous timestamps to learn effective node representations. This method first utilizes the historical information obtained from the network snapshots at past timestamps to create a spatial-temporal trajectory graph. Then it uses random walks with restart on the graph at a given timestamp as well as on graphs from past timestamps to capture the spatial-temporal behavior of nodes. The experiments on three real-world temporal network datasets show that the proposed learning method is valid. It also demonstrates the advantages of the method compared with both the state-of-the-art embedding techniques and several existing baseline methods.