Network Representation Learning Algorithm Integrating Local and Global Topology Information

ABSTRACT Network representation learning can map high‐dimensional sparse network data into a low‐dimensional vector space while preserving as much of the topology and node information as feasible. Network representation learning methods for obtaining network topology information have emerged in recent years. These methods do, however, not completely and adequately take into account the local or global information network topologies. To address this problem, we propose a novel network representation learning method. This method designs the local influence‐based random walk strategy and global similarity importance to separately capture the local and global topological information, aiming to improve representation learning's ability to learn network topological structure. On the local level, a new random walk metric named “local influence” is designed, which is constructed by fusing the structural entropy and degree importance of neighboring nodes of the target node. This metric is utilized to guide the random walk process, enabling the capture of local topological structure information in the network. On the global level, a global similarity importance‐based algorithm is proposed to filter out the global similarity importance nodes that serve as the global backbone of the network. These nodes are selected based on their global structural relevance, and their representations are learned from constructed subgraphs to encode the global topological properties in the network. Finally, the topological information obtained from both local and global levels is combined to enhance the overall node representations of the network. We conduct extensive experiments on five real‐world network datasets which demonstrate the effectiveness of the proposed approach.