Sharpening deep graph clustering via diverse bellwethers

Deep graph clustering has attracted increasing attention in data analysis recently, which leverages the topology structure and attributes of graph to divide nodes into different groups. Most existing deep graph clustering models, however, have compromised performance due to a lack of discriminative representation learning and adequate support for learning diverse clusters. To address these issues, we proposed a Diversity-promoting Deep Graph Clustering (DDGC) model that attains the two essential clustering principles of minimizing the intra-cluster variance while maximizing the inter-cluster variance. Specifically, DDGC iteratively optimizes the node representations and cluster centroids. First, DDGC maximizes the log-likelihood of node representations to obtain cluster centroids, which are subjected to a differentiable diversity regularization term to force the separation among clusters and thus increase inter-cluster variances. Moreover, a minimum entropy-based clustering loss is proposed to sharpen the clustering assignment distributions in order to produce compact clusters, thereby reducing intra-cluster variances. Extensive experimental results demonstrate that DDGC achieves state-of-the-art clustering performance and verifies the effectiveness of each component on common real-world datasets. Experiments also verify that DDGC can learn discriminative node representations and alleviate the over-smoothing issue.