l p -norm Multiple Kernel Learning with Diversity of Classes

Dayin Zhang, Hui Xue

Abstract: Multiple Kernel Learning (MKL) can learn an appropriate kernel combination from multiple base kernels for classification problems. It is often used to handle binary problems. However, multi-class problems appear in many real applications. In this paper, we propose a novel model, l p -norm multiple kernel learning with diversity of classes (LMKLDC), for the multi-class multiple kernel learning problem. LMKL-DC focuses on diversity of classes and aims to learn different kernel combinations for different classes to enhance the flexibility of our model. LMKLDC also utilizes l p -norm (0 < p ≤ 1) to promote the sparsity. However, LMKLDC boils down to a non-convex optimization problem when 0 < p < 1. In virtue of the constrained concave convex procedure (CCCP), we convert the non-convex optimization problem into a convex one and present a two-stage optimization algorithm. Experimental results on several datasets show our model selects fewer kernels and improves the classification accuracy.