Community detection is a fundamental problem in network analysis. In recent years, network representation learning has been leveraged to help the detection of potential communities. However, representation learning and community detection in these studies are usually optimized independently. In addition, they also have a strong assumption that the number of communities is prespecified before detection. To address those problems, we propose a joint representation learning and community detection (JRC) method. To allow the learned representations to be community-aware, JRC introduces two additional terms to the loss function for clustering. One is designed to getting node representations as close as possible to the exemplar of the community. Another is set up as a smoothness regularization to enforce the community properties of high cohesion and low coupling. Experimental results on multiple real-world network datasets show that JRC improves the quality of community detection and outperforms many competitive baselines.
展开▼
