As a promising clustering method, graph-based clustering converts the input data to a graph and regards the clustering as a graph partition problem. However, traditional graph clustering methods usually suffer from two main limitations: i), graph clustering is a feed-forward process, and cannot make use of the information from clustering result, which is more discriminative than the original graph; and ii), once the graph is constructed, the clustering process is no longer related to the input data, which may neglect rich information of raw features. To solve the above defects, we propose to learn the similarity graph adaptively, which compromises the information from the raw features, the initial graph and the clustering result. And thus, the proposed model is naturally cast as a joint model to learn the graph and generate the clustering result simultaneously, which is further efficiently solved with convergence theoretically guaranteed. The advantage of the proposed model is demonstrated by comparing with 19 state of-the-art clustering methods on 10 datasets with 4 clustering metrics.
View on arXiv