OCDIB: An Information Bottleneck-Guided Approach for Overlapping Community Detection

Community detection, aiming to group nodes in a graph with strong inner-connection, has attracted increasing research attention. However, existing methods for community detection assume the communities are non-overlapping, which are rare in real-world scenarios. For the overlapping community detection (OCD) task, each community might share overlapping nodes, and thus is difficult to train discriminative node representations to simultaneously distinguish different communities. To address this issue, we propose a novel approach for Overlapping Community Detection under the guidance of Information Bottleneck (OCDIB). Specifically, we separately model the distribution of communities at different latent spaces, from which we can sample the community-specific representation. In this way, we can better characterize the feature of each community. Then, to obtain salient representations, we propose a community-oriented information bottleneck. It aims to reduce noise in representations, while preserving the key information to express the target community. Last, we present a modularity-based loss function to guide the OCDIB to focus on the community-related structure and obtain high-quality community detection results. We have performed comprehensive experiments on both synthetic and real-world datasets, and the promising experimental results further demonstrate the effectiveness of our proposed approach w.r.t. a number of evaluation criteria.