package jsat.classifiers.boosting;
import java.util.List;
import java.util.ArrayList;
import jsat.classifiers.ClassificationDataSet;
import jsat.regression.Regressor;
import static jsat.utils.SystemInfo.*;
import jsat.utils.concurrent.ParallelUtils;
/**
* An extension to the original LogitBoost algorithm for parallel training.
* This comes at an increase in classification time.
*
* Note: LogitBoost is a semi unstable algorithm, and this method does in fact increase the instability.
* In most cases, similar classification results will be obtained, however - under some circumstances
* the performance may be significantly degraded. Especially if there is insufficient data to distribute
* for parallel computation. The results for LogitBoost seem to be over stated in the original paper.
*
* See: Scalable and Parallel Boosting with MapReduce, Indranil Palit and Chandan K. Reddy, IEEE Transactions on Knowledge and Data Engineering
* @author Edward Raff
*/
public class LogitBoostPL extends LogitBoost
{
private static final long serialVersionUID = -7932049860430324903L;
public LogitBoostPL(Regressor baseLearner, int M)
{
super(baseLearner, M);
}
public LogitBoostPL(int M)
{
super(M);
}
@Override
public void train(ClassificationDataSet dataSet, boolean parallel)
{
/*
* Implementation Note:
* In the original paper, we sort the weak hypotheses of the LogBoost workers
* by their unweighted accuracy. Then each regressor is averaged in its sorted
* group index. However, if we have M worked, each merged regressor is the sum
* of the parts divided by M, ie: the average. Applythis to the whole of the
* data set, we get the same result as if we add all regressors to the data
* set, getting M*Iteration hypothesis with the sume scaled by 1/(M)
* isntead of Iteration hypothesis scaled by 1/2
*
* Applied this, we can simplify the implementation and avoid M sortings
*
*/
List subSets = dataSet.cvSet(LogicalCores);
this.baseLearners = new ArrayList<>(LogicalCores * getMaxIterations());
ParallelUtils.streamP(subSets.stream(), parallel).forEach((subSet)->
{
LogitBoost boost = new LogitBoost(baseLearner.clone(), getMaxIterations());
boost.train(subSet);
for(Regressor r : boost.baseLearners)
baseLearners.add(r);
});
this.fScaleConstant = 1.0;
if(parallel)
this.fScaleConstant /= LogicalCores;
}
@Override
public LogitBoostPL clone()
{
LogitBoostPL clone = new LogitBoostPL(getMaxIterations());
clone.setzMax(getzMax());
if(this.baseLearner != null)
clone.baseLearner = this.baseLearner.clone();
if(this.baseLearners != null)
{
clone.baseLearners = new ArrayList<>(this.baseLearners.size());
for(Regressor r : baseLearners)
clone.baseLearners.add(r.clone());
}
return clone;
}
}
Ads help maintain this website.