Abstract
Community detection aims to identify a collection of closely connected nodes in a network, which is an important task in network analysis. In recent years, network embedding techniques based on random walks have been widely applied to community detection tasks. However, such methods suffer from numerous issues, such as ignoring node attributes and requiring the tedious process of manually setting random walk parameters. To address these issues, this paper proposes a network embedding method based on community-aware biased random walk for community detection in attributed networks. The method jointly leverages both the topological similarity and attribute similarity of nodes to guide the random walk process. When the random walk reaches the community boundary, the walker reduces the ambiguity of communities by enhancing the probability of walking toward the interior of the community. Compared with assigning unified walking parameters to all nodes, we adaptively set the number and length of random walks based on node degrees and the number of attributes. This strategy avoids the problem of oversampling nodes with low degrees and fewer attributes. Finally, we learn the node embeddings from the random walk sequences using the skip-gram model, and employ the clustering algorithm to obtain the community detection results. The effectiveness of the proposed algorithm is verified by conducting experiments on real-world and synthetic datasets.