An `sklearn` implementation of a feature selection procedure using p-value thresholds for each feature.

select_fisher.py

import sklearn
import fisher

class SelectPValueUnderThreshold(sklearn.base.TransformerMixin):
    """Select features with Fisher p-value under a certain threshold.
    """

    def __init__(self, threshold=1.0):
        self.threshold = threshold
        self.features_ = None
        self.pvalues_ = None

    def fit(self, X, y):
        self.pvalues_ = numpy.zeros(X.shape[1], dtype=numpy.float32)
        for feature in range(X.shape[1]):
            x = X[:,feature]
            result = fisher.pvalue(
                ((x > 0) & (y == 1)).sum(),
                ((x == 0) & (y == 1)).sum(),
                ((x > 0) & (y == 0)).sum(),
                ((x == 0) & (y == 0)).sum(),
            )
            self.pvalues_[feature] = result.two_tail
        self.indices_ = numpy.where(self.pvalues_ < self.threshold)[0]
        return self

    def transform(self, X):
        if self.indices_ is None:
            raise sklearn.exceptions.NotFittedError("model was not fitted")
        if X.shape[1] != self.pvalues_.shape[0]:
            raise ValueError(f"X has {X.shape[1]} features, but SelectPValueUnderThreshold is expecting {self.pvalues_.shape[0]} features as input.")
        return X[:, self.indices_]