String kernels (coming soon)

The design of kernelmethods allows for easy implementation to accept any input data types, as that only matters in the definition of the kernel function which are modularized out of the remaining parts of library.

For example, to implement mismatch kernels for protein sequence data, it is simply a matter of implementing k-mer feature extraction tool e.g. kmer_spectrum(). Then a simple mismatch kernel can easily be defined in few lines via:

from kernelmethods.base import KernelFromCallable
from kernelmethods.strings import kmer_spectrum # Note: NotImplemented
import numpy as np

def kmer_similarity(string_one, string_two):

    return np.dot(kmer_spectrum(string_one), kmer_spectrum(string_two))

mismatch_kernel = KernelFromCallable(kmer_similarity)

Or one could also turn this into a full-fledged StringKernel or MismatchStringKernel classes, while reusing a lot of boilerplate from kernelmethods.base mimicking numerical or categorical kernels already defined.

Once a MismatchStringKernel is available, we can leverage the full functionality of the kernelmethods to the user domain/application.