Soft-clustering-based local multiple kernel learning algorithm for classification

Abstract

Local multiple kernel learning is a promising strategy because it could learn a sample-specific composite kernel according to the characteristic of samples. However, these methods are insufficient to describe the sample diversity and correlation, which leads to the classifier less reliable. In this paper, we propose a soft-clustering-based local multiple kernel learning algorithm to tackle the issues above. In the proposed algorithm, there is a fuzzy clustering preprocessing for the training data and then the kernel weights are calculated on the groups. We use an alternative optimization method to learn the kernel weights and support vector coefficients. The final combination weights of kernels are determined by the kernel weights of clusters and the probability of samples falling into the clusters. Therefore, our method is actually a sample-based LMKL method with a soft constraint on the kernel weights. This constraint is actually the representation of the correlation of samples. The experiments on synthetic dataset indicate the kernel weights solved by our algorithm are better suitable for the characteristics of the dataset. Then a series of experiments verify the improvement on classification accuracies.