Multiple Kernel K-means Clustering : A Self-weighted Learning Approach

The k-means algorithm is one of the most widely used methods for data clustering. However, traditional k-means can only be applied in the original feature space, limited to handle nonlinear separable clusters. To address this issue, kernel k-means is proposed to map low-dimensional linearly inseparable data to a high-dimensional space, and the mapped data are expected to be linearly separable. However, same as other kernel methods, the success of kernel k-means is highly depended on the choice of the kernel function. Unfortunately, it is difficult to determine which kernel is the most suitable for a specific task in practical applications. Naturally, it is thought that multiple kernel learning can perform better than a single kernel, but it is inevitable to introduce multiple weight parameters. In this paper, we propose a novel multiple kernel k-means method, which adopts a self-weighted learning approach. Unlike other multiple kernel k-means methods in the literature, our method can learn a consensus kernel and automatically assign an optimal weight to each kernel without introducing additional parameters. Extensive experiments on eight benchmark datasets compared with several state-of-the-art multiple kernel clustering methods demonstrate our new multiple kernel k-means method can achieve comparable performance and can be used more practically.