Multiple Kernel Clustering with Direct Consensus Graph Learning

Multiple kernel graph-based clustering (MKGC) has achieved impressive experimental results, primarily due to the superiority of multiple kernel learning (MKL) and the outstanding performance of graph-based clustering. However, many present MKGC methods face the following two disadvantages that pose challenges for further improving clustering performance: (1) these methods always rely onMKL to learn a consensus kernel from multiple base kernels, which may lose some important graph information since graph learning is the key to graph-based clustering, not kernel learning; (2) these methods perform affinity graph learning and subsequent graph-based clustering in two separate steps, which may not be optimal for clustering tasks. To tackle these problems, this paper proposes a new MKGC method for multiple kernel clustering. By directly learning a consensus affinity graph rather than a consensus kernel from multiple base kernels, the important graph information can be preserved. Moreover, by utilizing rank constraint, the cluster indicators are obtained directly without performing the k-means clustering and any graph cut technique. Extensive experiments on benchmark datasets demonstrate the superiority of the proposed method.