Cluster center consistency guided sampling learning for multiple kernel clustering

Abstract

Multiple kernel clustering (MKC) is popular in processing non-linear data over the past years. The main challenge is that the kernel matrix with the size n×n leads to high memory and computational consumption, where n denotes the number of samples. Although the widely used static anchor sampling can mitigate such a challenging considerably, how to dynamically select anchor points with a learning manner from kernel matrix is a difficult problem. To address the issue, this paper proposes a novel method dubbed as cluster center consistency guided sampling learning (3CSL) for multiple kernel clustering (3CSL-MKC). Specifically, by taking the cluster center consistency between the original partial kernel data and anchor points into consideration, 3CSL-MKC learns the shared anchor sampling matrix gradually. With the help of high-quality anchors, the essential clustering information of each kernel partition can be transformed largely into a concentrated low-dimensional representation matrix. Meanwhile, based on dictionary learning, 3CSL-MKC fuses these candidate representation matrixes to produce the resulting consensus representation, such that the clustering assignments can be directly obtained relying on simple k-means. A large number of experiments are conducted on different multiple kernel datasets to verify the effectiveness and efficiency of the proposed method.