Multiple Kernel Learning by Gap Maximization between Classes in RKHS

Abstract

By implicitly maximizing the gap between classes in the reproducing kernel Hilbert space (RKHS), a multiple kernel learning (MKL) is formulated as a linear programming in this paper. For each sample, my method tries to enforce the distance between intra-class and inter-class samples in RKHS to be as distant as possible. n my method, each training sample imposes at most r constraints for the linear programming where r is the number of different kernel types. Unlike previous methods of multiple kernel learning, the proposed method does not need a large amount of computations. My method is compared with various methods of MKL to prove its efficiency of finding a good kernel mixture parameter.