GROOT forest

The GrootRandomForestClassifier class uses bootstrap aggregation and partially random feature selection to train an ensemble of GrootTreeClassifiers. On datasets with many features, a GrootRandomForestClassifier might perform better than a GrootTreeClassifier as it is not limited in the number of features it can use by a maximum size.

Example:

from sklearn.datasets import make_moons
X, y = make_moons(random_state=1)

from groot.model import GrootRandomForestClassifier
forest = GrootRandomForestClassifier(attack_model=[0.1, 0.1], random_state=1)
forest.fit(X, y)
print(forest.score(X, y))

1.0

groot.model.GrootRandomForestClassifier (BaseGrootRandomForest, ClassifierMixin)

A robust random forest for binary classification.

__init__(self, n_estimators=100, max_depth=None, max_features='sqrt', min_samples_split=2, min_samples_leaf=1, robust_weight=1.0, attack_model=None, one_adversarial_class=False, verbose=False, chen_heuristic=False, max_samples=None, n_jobs=None, compile=True, random_state=None) special

Parameters:

Name	Type	Description	Default
n_estimators	int	The number of decision trees to fit in the forest.	100
max_depth	int	The maximum depth for the decision trees once fitted.	None
max_features	int or {"sqrt", "log2", None}	The number of features to consider while making each split, if None then all features are considered.	'sqrt'
min_samples_split	int	The minimum number of samples required to split a tree node.	2
min_samples_leaf	int	The minimum number of samples required to make a tree leaf.	1
robust_weight	float	The ratio of samples that are actually moved by an adversary.	1.0
attack_model	array-like of shape (n_features,)	Attacker capabilities for perturbing X. The attack model needs to describe for every feature in which way it can be perturbed.	None
one_adversarial_class	bool	Whether one class (malicious, 1) perturbs their samples or if both classes (benign and malicious, 0 and 1) do so.	False
verbose	bool	Whether to print fitting progress on screen.	False
chen_heuristic	bool	Whether to use the heuristic for the adversarial Gini impurity from Chen et al. (2019) instead of GROOT's adversarial Gini impurity.	False
max_samples	float	The fraction of samples to draw from X to train each decision tree. If None (default), then draw X.shape[0] samples.	None
n_jobs	int	The number of jobs to run in parallel when fitting trees. See joblib.	None
compile	bool	Whether to compile decision trees for faster predictions.	True
random_state	int	Controls the sampling of the features to consider when looking for the best split at each node.	None

Attributes:

Name	Type	Description
estimators_	list of GrootTree	The collection of fitted sub-estimators.
n_samples_	int	The number of samples when fit is performed.
n_features_	int	The number of features when fit is performed.

predict(self, X)

Predict the classes of the input samples X.

The predicted class is the rounded average of the class labels in each predicted leaf.

Parameters:

Name	Type	Description	Default
X	array-like of shape (n_samples, n_features)	The input samples to predict.	required

Returns:

Type	Description
array-like of shape (n_samples,)	The predicted class labels

predict_proba(self, X)

Predict class probabilities of the input samples X.

The class probability is the average of the probabilities predicted by each decision tree. The probability prediction of each tree is the fraction of samples of the same class in the leaf.

Parameters:

Name	Type	Description	Default
X	array-like of shape (n_samples, n_features)	The input samples to predict.	required

Returns:

Type	Description
array of shape (n_samples,)	The probability for each input sample of being malicious.