Feature selection by mutual information
A=MUTINF(C,H) returns a mutinf object initialized with hyperparameters H given a classifier C.
Performs feature selection by means of the mutual information between attributes and the target.
If there is a given number of features to select, this number of features is selected according to
the ranking based upon the mutatul information.
If no number of features to select is given, a probalistic forward selection is used:
A feature is selected, if P(I > epsilon) >= 95 , where I is the mutual information between the attribute
and the target. (c.f. Zaffalon, Hutter, Robust Feature Selection by Mutual Information Distributions)
Here epsilon is set to the mutual information that exists between a normally distributed random feature
and the target.
Hyperparameters, and their defaults
feat= -- number of features to be selected
method='regression' -- use feature selection for regression or classification
c -- learning algorithm (e.g. svm)
rank -- the ranking of the features
child -- learning algorithm (e.g. svm)
train, test, get_w
Reference : Robust Feature Selection by Mutual Information Distributions |
Author : Marco Zaalon and Marcus Hutter |