Mismatch string kernel

This module can be used to vectorize strings and compute kernel between them.

A Python3 implementation of the mismatch kernel described in the publication below:

%0 Journal Article
%T Mismatch string kernels for discriminative protein classification
%A Leslie, Christina S
%A Eskin, Eleazar
%A Cohen, Adiel
%A Weston, Jason
%A Noble, William Stafford
%J Bioinformatics
%V 20
%N 4
%P 467-476
%@ 1460-2059
%D 2004
%I Oxford University Press
%U https://doi.org/10.1093/bioinformatics/btg431

Usage

To understand the technicalities of what this kernel does please refer to the article above.

Initializing the kernel

First you have to define an alphabet from which the k-mers will be generated, the length k of the k-mers and m the maximum number of mismatches between mers, for example:

ALPHABET = ['a', 'b', 'c', 'd', 'e', 'f', 'g', 'h', 'i', 'j', 'k', 'l', 'm',
            'n', 'o', 'p', 'q', 'r', 's', 't', 'u', 'v', 'w', 'x', 'y', 'z', ' ']
k = 3
m = 1

Then you can create a MismatchKernel object with such parameters:

from mismatch_kernel import MismatchKernel

mk = MismatchKernel(ALPHABET, k, m)

Mapping a string to (k-m) feature space

You can use the vectorize(x) function to map a string x to the (k-m) feature space.

Note that the alphabet is in general case sensitive, so if your strings needs to be case sensitive (i.e. "string" != "StRiNg"), your alphabet should contain both uppercase and lowercase letters. Also this will much increase computational time because the k-mer feature space has dimension #(ALPHABET)^k; same thing goes if you need to distinguish punctuation, for example in the alphabet above the strings will be different based on the spaces they contain (i.e. "space" != "spa ce"). In general the strings you pass to this module functions will be normalized, i.e. every character not in the alphabet will be removed. For example if you call vectorize("String") after defining the above alphabet you are actually vectorizing "tring", so you should have called vectorize("String".lower()) instead.

The vectorize(x) function returns a tuple (x_norm, dok) where x_norm is the actual string that has been vectorized (i.e. x normalized) so you can check if that's what you actually wanted to vectorize, and dok is the vector in DOK (dictionary of keys) format (because the vectors are generally sparse), so it will be a dictionary like {2: 1, 3: 1, 14: 1, 17: 2, 30: 1, 41: 1, ...} meaning that the vector has non-zero values only at the position of the dictionary keys, i.e. [0, 0, 1, 1, 0, ..., 0, 1, 0, 0, 2, ...]. You can push x_norm in a dictionary along with the vector so you don't have to vectorize it again, this is what the get_kernel() function actually does.

Example

x_norm, vect = mk.mismatch_tree.vectorize("doc. Frankenstein".lower())
print("{} -> {}".format(x_norm, vect))

> doc frankenstein -> {10: 1, 13: 1, 37: 1, 64: 1, ...}

Calculating the kernel between two strings

You can use the get_kernel(x1, x2) function to get the kernel between x1 and x2, the kernel varies between 0 and 1, the more similar the two strings the greater it will be (1 if the strings are equal). The function will automatically normalize and vectorize the two strings to compute the kernel.

Example

ker = mk.get_kernel("doc. Frankenstein".lower(), "doc. Drunkenstein".lower())
print(ker)

> 0.7500011542039571

Using or supplying already calculated mismatch vectors and kernels

The get_kernel function will save in the MismatchKernel object the mismatch vectors of every string it vectorizes in the MISMATCH_VECTORS attribute, that is a dictionary that stores strings as keys and the corresponding vector as values (i.e. {'doc frankenstein': {10: 1, 13: 1, 37: 1, 64: 1, ...}, 'doc drunkenstein': {80: 1, 98: 1, 116: 1, 121: 1, ...}}) so if you call next mk.get_kernel("doc drunkenstein", "doc nykterstein" it won't vectorize again "doc drunkenstein".

Likewise every calculated kernel will be stored in the KERNEL_MATRIX attribute, that is a dictionary that stores strings as keys and another dictionary with strings as keys and the kernel value between the two keys as values (i.e. {'doc frankenstein': {'doc drunkenstein': 0.7500011542039571, 'doc nykterstein': 0.5041614599291009}}). If you have to calculate the kernel for a batch of strings you can call get_kernel from the same MismatchKernel object so the strings for which the mismatch vector or the kernel have already been calculated won't be calculated again.

If you already have one or both of these dictionaries, for example if you pickled the MISMATCH_VECTORS and KERNEL_MATRIX attributes from a previous run, you can pass them to the MismatchKernel constructor:

mk = MismatchKernel(ALPHABET, k, m, vectors_dict, kernels_dict)