A pre-selecting base kernel method in multiple kernel learning

The pre-defined base kernel greatly affects the performance of multiple kernel learning (MKL), but selecting the pre-defined base kernel still has no theoretical guidance. In practice, it is very difficult to select a set of appropriate base kernels without prior knowledge. In this paper, we propose a general strategy to pre-select a reasonable set of base kernels before the optimization process of MKL solvers. This strategy is based on the combination of minimal redundancy maximal relevance criteria and kernel target alignment (MRMRKA). First, we determine some candidate kernels while maintaining diversity of information; second, a set of base kernels with high discriminative ability and large diversity are selected using the MRMRKA method. These pre-selected base kernels will be used in the optimization process of the existing MKL solvers to generate better results. The experiments conducted on UCI and 15-scene datasets show that the performance of MKL is improved with the proposed pre-selected base kernel strategy.