RandomForestClassifier

A forest of continuous-feature classification trees: each is fitted to a bootstrap sample of the rows and considers a sample of the features at every node, and the scores are averaged. RandomForest above is built on the binary-feature DecisionTree, so continuous data must be binned before it can be used. This is built on DecisionTreeClassifier and splits on thresholds directly. Both kinds of sampling matter, for different reasons. Bootstrapping the rows and sampling the features decorrelates the trees, which is what makes the average better than any single tree. Sampling the features is also the cheapest thing available on build time: the split search dominates and is linear in the features tried, so the default of sqrt(features) is roughly that speedup over trying all of them. Deterministic: the same seed rebuilds the same forest, so a result can be reproduced rather than merely repeated.

Example

forest := RandomForestClassifier->New(100, 8, 5, 42);
forest->Fit(x_train, y_train);
scores := forest->PredictScore(x_test);
recall := Metrics->RecallAtFpr(scores, y_test, 0.001);

Operations

• New
• Fit
• GetTreeCount
• IsFitted
• Predict
• PredictScore

method : public : Fit(X:Float[,], y:Bool[]) ~ Bool

method : public : GetTreeCount() ~ Int

method : public : IsFitted() ~ Bool

New(num_trees:Int, max_depth:Int, min_samples:Int, seed:Int)

New(num_trees:Int, max_depth:Int, min_samples:Int, max_features:Int, seed:Int)

method : public : Predict(X:Float[,]) ~ Bool[]

method : public : PredictScore(X:Float[,]) ~ Float[]