import numpy as np
from sklearn.linear_model import LogisticRegression
[docs]class LogisticRegressionWithThreshold:
"""
A modified Logistic Regression classifier that will filter calls by a custom threshold, instead of the default 0.5. This allows for control of the precision-recall tradeoff, e.g. false positives versus false negatives.
:ivar clf: The underlying LogisticRegression classifier
:ivar threshold: Threshold to use, should be between 0 and 1
"""
[docs] def __init__(self,threshold=0.5):
"""
Set up a Logistic Regression classifier that can use a different threshold for predictions and thereby be more lenient (lower threshold, false positives increase, false negatives decrease) or more conservative (higher threshold, false positives decrease, false negative increase).
:param threshold: Threshold to use, should be between 0 and 1
:type threshold: float
"""
assert threshold >= 0 and threshold <= 1, "Threshold must be between 0 and 1"
self.clf = LogisticRegression(class_weight='balanced',random_state=1,solver='liblinear',multi_class='ovr')
self.threshold = threshold
[docs] def fit(self,X,Y):
"""
Train the classifier using the associated matrix X and classes Y. Class zero should represent no associated class.
:param X: Training vector
:param Y: Associated class for each row of X
:type X: sparse matrix
:type Y: matrix
"""
self.clf.fit(X,Y)
self.classes_ = self.clf.classes_
[docs] def predict(self,X):
"""
Make predictions for the class of each row in X. Class zero should represent no prediction.
:param X: Testing vector
:type X: sparse matrix
:return: Predictions of classes for each row in X
:rtype: matrix
"""
probs = self.clf.predict_proba(X)
# Ignore probabilities that fall below our threshold
probs[probs<self.threshold] = -1.0
# Make sure that the zero class is only select if all other options are below the threshold
probs[:,0] = -0.5
# And get the highest probability for each row
predictions = np.argmax(probs,axis=1)
return predictions
[docs] def predict_proba(self,X):
"""
Calculate probabilities for the class of each row in X. Class zero should represent no prediction.
Returns a matrix of probabilities
:param X: Testing vector
:type X: sparse matrix
:return: Probabilities of classes for each row in X
:rtype: matrix
"""
return self.clf.predict_proba(X)