A novel random walk strategy for network embedding using community aware and node influence biasing

Abstract

Network embedding (NE), a central issue for graph deep learning research, intends to represent nodes, edges, sub-graphs, and other information into a low-dimensional vector space. A highly accurate NE algorithm benefits significantly in developing AI applications related to social platforms such as node classification and link prediction. Existing studies mainly focus on preserving the local and global information of nodes. The propagation of bias information is not fully considered according to the desires of the nodes. Incorporating community aware and node influence into the random walk process for NE issues are not addressed. This work proposes a highly accurate NE algorithm, namely Community Aware and Node Influence Biased Embedding (CNBE), to deal with the NE problem of social networks. CNBE recognizes the community information and the random walk with node influence bias to preserve the global structure of nodes and local neighborhood information of reserved nodes. By random walk sampling of predefined rules in CNBE, we can get the sequence of nodes that reserved the local and global neighborhood node information. Then the SkipGram model is applied to learn the low-dimensional vector representation of nodes. Experiments on three real-world networks indicate that our proposed algorithm method outperforms three state-of-the-art network embedding approaches.