Gradient Boosted Decision Trees (GBDT)

This library provides support for gradient boosted decision trees.

Basic Usage

For an example of how to use this library, see the Gradient Boosted Decision Trees tutorial.

Interface

To create an instance of the GBDT module, use

structure C = CartReal(structure D = DecisionTreeReal)
structure Gbdt = Gbdt(structure CART = C)

and then prefix the types and functions below with Gbdt..

Types

• structure DT : DECISION_TREE
• structure CART : CART_REAL
  • The GBDT libraries are parameterized by the decision trees and CART libraries, which provide the label and feature types used. See the decision tree and CART documentation for information on those types.
• type t
  • The type of an ensemble of decision trees.

Methods

• val forward: t * DT.features -> DT.label
• val error: t * (DT.features * DT.label) list -> real
• val train: (DT.features * DT.label) list * real * int -> t
• val toString: t -> string

Method Overview

• forward (gbdt, ft)
  • Uses the ensemble gbdt to predict the label for the feature vector ft.
• error (gbdt, data)
  • Returns the error of the ensemble gbdt on the labeled data.
• train (data, learningRate, numEpochs)
  • Trains ensemble of decision trees on the given data, using the learning rate, for the given number of epochs.
• toString gbdt
  • Converts an ensemble to a string.