Multiple Kernel Clustering with Kernel k-Means Coupled Graph Tensor Learning

Kernel k-means (KKM) and spectral clustering (SC) are two basic methods used for multiple kernel (MKC), which have both been widely to identify clusters that non-linearly separable. However, of them their own shortcomings: 1) the KKM-based usually focus on learning a discrete indicator matrix via combined consensus kernel, but cannot exploit high-order affinities all pre-defined base kernels; 2) SC-based require robust meaningful affinity graph in space as input order form with desired structure. In this paper, novel method, coupled tensor (KCGT), is proposed graciously couple KKM SC seizing merits evading demerits simultaneously. specific, we innovatively develop new paradigm by leveraging an explicit theoretical connection between graph, such propagated from enjoys valuable block diagonal sparse property. Then, using paradigm, kernels can produce candidate graphs, stacked into low-rank capturing these graphs. After that, averaging frontal slices tensor, high-quality obtained. Extensive experiments shown superiority KCGT compared state-of-the-art MKC methods.