We introduce a family of positive definite kernels specifically designed for problems described by categorical information. The kernels are based on the comparison of the probability mass function of the variables and have a clear interpretation in terms of similarity computations between the modalities. We report experimental results on two different problems in the life sciences indicating that the proposed approach may markedly outperform standard kernels, so it can be used as a good alternative to other common kernel functions (at least for SVM classification) in order to obtain better accuracy.

Peer Reviewed