biasedclassifier 0.3.1

No project description provided

Biased Classifier

Biased Classifier

Current version: 0.3.0

Install

Directly from PyPi servers:

pip install biasedclassifier

Interface

Estimator's constructor:

BiasedClassifier(
    p=[0.0],
    unbiased_estimator=None,
    knn=None
)

where unbiased_estimator is the base estimator to use (and to biased towards critical set). We pass a k-NearestNeighbor object directly via the paramter knn.

Use

Example using Random Forests from scikit-learn.

Assume X, y is a training set with three classes and two heavily inbalanced classes. In this case, we'd like to bias two classifiers into these subsets. We've decided that 0.3 and 0.2 proportions are enough for the minority classes (from smaller up) and k=10 neighbors to collect for critical set. Our unbiased estimator will be a random forest of size 200.

from biasedclassifier import BiasedClassifier
from sklearn.neighbors import NearestNeighbors
from sklearn.ensemble import RandomForestClassifier

clf = BiasedClassifier(
    p=[0.3, 0.2], 
    unbiased_classifier=RandomForestClassifier(n_estimators=200), 
    knn=NearestNeighbors(n_neighbors=10)
)

# Train
clf.fit(X,y)

# Obtain probabilities for each class
prob = clf.predict_proba(X)

# Predicted values
y_pred = clf.predict(X)

# Average accuracy score
score = clf.score(X, y)

It is important to note that BiasedEstimator does not change the state of both objects unbiased_classifier and knn. Instead, it uses clones internally to do its operations.

Compatibility

This model is compatible with all of the capabilities offered by scikit-learn requiring get_params and score methods.