MKPLS approach: switching strategies for the non-linear multi-kernel PLSR

We present two strategies to determine the kernel switching order for the non-linear multi-kernel PLSR algorithm. The multi-kernel PLS (MKPLS) algorithm builds upon a one kernel PLSR which uses a kernel matrix to hold the inner products of the projection of the independent data set onto a feature space. After building a PLSR model, MKPLS deflates the kernel matrix so that only that part which cannot be predicted by the model remains. This remainder is projected onto a different feature space and a new PLSR model is built. The switching algorithms presented for this approach address two questions: which kernel should be used at each iteration and; how many factors should be extracted before switching to another kernel.