IsoRandomForestClassifier is a Random Forest classifier with globally guaranteed monotonicity and partial monotonicity capability (i.e. the ability to specify both monotone and non-monotone features). It extends scikit-learn's RandomForestClassifier and inherits all sci-kit learn capabilities (and obviously requires sci-kit learn). It is described in Chapter 6 of the PhD thesis 'High Accuracy Partially Monotone Ordinal Classification', UWA 2019.
First we define the monotone features, using the corresponding one-based X array column indices:
incr_feats=[6,9]
decr_feats=[1,8,13]
The specify the usual RF hyperparameters:
# Ensure you have a reasonable number of trees
n_estimators=200
mtry = 3
And initialise and solve the classifier using scikit-learn norms:
clf = isoensemble.IsoRandomForestClassifier(n_estimators=n_estimators,
max_features=mtry,
incr_feats=incr_feats,
decr_feats=decr_feats)
clf.fit(X, y)
y_pred = clf.predict(X)
Of course usually the above will be embedded in some estimate of generalisation error such as out-of-box (oob) score or cross-validation.
BSD 3 Clause, Copyright (c) 2017, Christopher Bartley All rights reserved.