Package ciir.umass.edu.learning.tree

Class LambdaMART

java.lang.Object

ciir.umass.edu.learning.Ranker

ciir.umass.edu.learning.tree.LambdaMART

Direct Known Subclasses:

MART

public class LambdaMART
extends Ranker

Author:

vdang This class implements LambdaMART. Q. Wu, C.J.C. Burges, K. Svore and J. Gao. Adapting Boosting for Information Retrieval Measures. Journal of Information Retrieval, 2007.

Field Summary

Fields

Modifier and Type	Field	Description
protected int	bestModelOnValidation
protected Ensemble	ensemble
static int	gcCycle
protected FeatureHistogram	hist
static float	learningRate
protected DataPoint[]	martSamples
static int	minLeafSupport
protected double[]	modelScores
protected double[][]	modelScoresOnValidation
static int	nRoundToStopEarly
static int	nThreshold
static int	nTreeLeaves
static int	nTrees
protected double[]	pseudoResponses
protected int[][]	sortedIdx
protected float[][]	thresholds
protected double[]	weights

Fields inherited from class ciir.umass.edu.learning.Ranker

bestScoreOnValidationData, features, samples, scoreOnTrainingData, scorer, validationSamples, verbose

Constructor Summary

Constructors

Constructor	Description
LambdaMART()
LambdaMART(java.util.List<RankList> samples, int[] features, MetricScorer scorer)

Method Summary

Modifier and Type	Method	Description
protected float	computeModelScoreOnTraining()
protected float	computeModelScoreOnTraining(int start, int end, int current)
protected float	computeModelScoreOnValidation()
protected float	computeModelScoreOnValidation(int start, int end)
protected void	computePseudoResponses()
protected void	computePseudoResponses(int start, int end, int current)
Ranker	createNew()
double	eval(DataPoint dp)
Ensemble	getEnsemble()
void	init()	HAVE TO BE OVER-RIDDEN IN SUB-CLASSES
void	learn()
void	loadFromString(java.lang.String fullText)
java.lang.String	model()
java.lang.String	name()
void	printParameters()
protected RankList	rank(int rankListIndex, int current)	This function is equivalent to the inherited function rank(...), but it uses the cached model's outputs instead of computing them from scratch.
protected void	sortSamplesByFeature(int fStart, int fEnd)
protected int[]	sortSamplesByFeature(DataPoint[] samples, int fid)
java.lang.String	toString()
protected void	updateTreeOutput(RegressionTree rt)

Methods inherited from class ciir.umass.edu.learning.Ranker

copy, getFeatures, getScoreOnTrainingData, getScoreOnValidationData, PRINT, PRINT, PRINT_MEMORY_USAGE, PRINTLN, PRINTLN, PRINTTIME, rank, rank, save, setFeatures, setMetricScorer, setTrainingSet, setValidationSet

Methods inherited from class java.lang.Object

clone, equals, finalize, getClass, hashCode, notify, notifyAll, wait, wait, wait

Field Detail

nTrees

public static int nTrees

learningRate

public static float learningRate

nThreshold

public static int nThreshold

nRoundToStopEarly

public static int nRoundToStopEarly

nTreeLeaves

public static int nTreeLeaves

minLeafSupport

public static int minLeafSupport

gcCycle

public static int gcCycle

thresholds

protected float[][] thresholds

ensemble

protected Ensemble ensemble

modelScores

protected double[] modelScores

modelScoresOnValidation

protected double[][] modelScoresOnValidation

bestModelOnValidation

protected int bestModelOnValidation

martSamples

protected DataPoint[] martSamples

sortedIdx

protected int[][] sortedIdx

hist

protected FeatureHistogram hist

pseudoResponses

protected double[] pseudoResponses

weights

protected double[] weights

Constructor Detail

LambdaMART

public LambdaMART()

LambdaMART

public LambdaMART(java.util.List<RankList> samples,
                  int[] features,
                  MetricScorer scorer)

Method Detail

init

public void init()

Description copied from class: Ranker

HAVE TO BE OVER-RIDDEN IN SUB-CLASSES

Specified by:

init in class Ranker

learn

public void learn()

Specified by:

learn in class Ranker

eval

public double eval(DataPoint dp)

Overrides:

eval in class Ranker

createNew

public Ranker createNew()

Specified by:

createNew in class Ranker

toString

public java.lang.String toString()

Specified by:

toString in class Ranker

model

public java.lang.String model()

Specified by:

model in class Ranker

loadFromString

public void loadFromString(java.lang.String fullText)

Specified by:

loadFromString in class Ranker

printParameters

public void printParameters()

Specified by:

printParameters in class Ranker

name

public java.lang.String name()

Specified by:

name in class Ranker

getEnsemble

public Ensemble getEnsemble()

computePseudoResponses

protected void computePseudoResponses()

computePseudoResponses

protected void computePseudoResponses(int start,
                                      int end,
                                      int current)

updateTreeOutput

protected void updateTreeOutput(RegressionTree rt)

sortSamplesByFeature

protected int[] sortSamplesByFeature(DataPoint[] samples,
                                     int fid)

rank

protected RankList rank(int rankListIndex,
                        int current)

This function is equivalent to the inherited function rank(...), but it uses the cached model's outputs instead of computing them from scratch.

Parameters:

rankListIndex -

current -

Returns:

computeModelScoreOnTraining

protected float computeModelScoreOnTraining()

computeModelScoreOnTraining

protected float computeModelScoreOnTraining(int start,
                                            int end,
                                            int current)

computeModelScoreOnValidation

protected float computeModelScoreOnValidation()

computeModelScoreOnValidation

protected float computeModelScoreOnValidation(int start,
                                              int end)

sortSamplesByFeature

protected void sortSamplesByFeature(int fStart,
                                    int fEnd)