Multiple kernel learning with hybrid kernel alignment maximization

Abstract

Two-stage multiple kernel learning (MKL) algorithms have been extensively researched in recent years due to their high efficiency and effectiveness. Previous works have attempted to optimize the combination coefficients by maximizing the centralized kernel alignment between the combined kernel and the ideal kernel. Though demonstrating previous promising performance, we observe that these algorithms may suffer from the approaching in calculating the alignment. In particular, we observe that the local information should be incorporated when computing the kernel alignment, which is beneficial to further improve the classification performance. To this end, we first define the local kernel alignment based on centralized kernel alignment. A new kernel alignment that combines the global and local information of base kernels is then developed. After that, we propose an alternative algorithm with proved convergence to identify the multiple kernel coefficients. Intensive experimental results show that the performance of the proposed algorithm is superior to those of existing MKL algorithms.