binaryprobabilisticsplitter.h

Go to the documentation of this file.

/*

Copyright (c) 2003, Cornell University
All rights reserved.

Redistribution and use in source and binary forms, with or without
modification, are permitted provided that the following conditions are met:

   - Redistributions of source code must retain the above copyright notice,
       this list of conditions and the following disclaimer.
   - Redistributions in binary form must reproduce the above copyright
       notice, this list of conditions and the following disclaimer in the
       documentation and/or other materials provided with the distribution.
   - Neither the name of Cornell University nor the names of its
       contributors may be used to endorse or promote products derived from
       this software without specific prior written permission.

THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF
THE POSSIBILITY OF SUCH DAMAGE.

*/

// -*- C++ -*-

#if !defined _CLUS_BINARYPROBABILISTICSPLITTER_H_
#define _CLUS_BINARYPROBABILISTICSPLITTER_H_

#include "general.h"
#include "statisticsgatherers.h"
#include "vec.h"

#ifdef DEBUG_PRINT
#include <iostream>
using namespace std;
#endif

using namespace TNT;

namespace CLUS
{
class BinaryProbabilisticSplitter
{
protected:
    /// the number of discrete and continuous split variables
    int dsplitDim, csplitDim;

    /// statistics about datapoints seen
    double mass0, mass; // sum of probabilities

    /// list of discrete domain sizes
    const Vector<int>& dDomainSize;
    
    /** Indicates on what variable this node splits on. */
    int SplitVariable;

    /// split for continuous variables
    double splitPoint, splitSTD; 

    Vector<double> splitSetProbability;

    /** Statistics for split decision. One for each attributeXdataset
        If an attribute already has the shift determined keep statistics
        about all datasets in the location for the first dataset
    */
    Vector<ProbabilisticBinomialStatistics>  discreteStatistics;
    Vector<NormalStatistics>  continuousStatistics;

public:
    BinaryProbabilisticSplitter(const Vector<int>& DDomainSize,int CsplitDim ):
            dsplitDim(DDomainSize.dim()-1), csplitDim(CsplitDim),
            dDomainSize(DDomainSize),
            discreteStatistics(0), continuousStatistics(0)
    {}

    bool GotNoData(void)
    {
        return mass==0.0;
    }

    int GetCSplitDim(void)
    {
        return csplitDim;
    }
    
    int GetDSplitDim(void)
    {
        return dsplitDim;
    }
    
    const Vector<int>& GetDDomainSize(void)
    {
        return dDomainSize;
    }

    /** Initializes the data structures used in split variable selection */
    void InitializeSplitStatistics(void)
    {
        mass0=mass=0.0;

        discreteStatistics.newsize(dsplitDim);
        for (int i=0; i<dsplitDim; i++)
        {
            discreteStatistics[i].ResetDomainSize(dDomainSize[i]);
        }

        continuousStatistics.newsize(csplitDim);

    }

    /** Computes the probability to take the left branch */
    double ProbabilityLeft(const int* Dvars, const double* Cvars)
    {
        if (SplitVariable <=-1)
        {
            // split on a continuous variable
            // the split variable is actually -SplitVariable-1
            int splitVar=-SplitVariable-1;
            // return the CDF of N(splitPoint,splitSTD)(current point)
            return PValueNormalDistribution(splitPoint,splitSTD,Cvars[splitVar]);
        }
        else
        {
            // split on a discrete variable
            int value=Dvars[SplitVariable];
            return splitSetProbability[value];
        }
    }

    void UpdateSplitStatistics( const int* Dvars, const double* Cvars,
                                int classLabel, double probability)
    {
        mass+=probability;
        if (classLabel==0)
            mass0+=probability;

        // update discrete statistics
        for (int i=0; i<dsplitDim; i++)
        {
            int value=Dvars[i];
            discreteStatistics[i].UpdateStatistics(value, classLabel, probability);
        }

        // update continuous statistics
        for (int i=0; i<csplitDim; i++)
        {
            double value=Cvars[i];
            continuousStatistics[i].UpdateStatistics(value, classLabel, probability);
        }
    }

    bool ComputeSplitVariable(void)
    {
        // make node a leaf if not enough data to take a split decision
        if (mass0<2.0 || mass-mass0<2.0)
        {
#ifdef DEBUG_PRINT
            cout << "Making the node a leaf" << endl;
#endif

            return false;
        }

        double maxgini=0.0;

        // go over the discrete attributes and find the best one
        for (int i=0; i<dsplitDim; i++)
        {
            double curr_gini=discreteStatistics[i].ComputeGiniGain();
#ifdef DEBUG_PRINT

            cout << "\tVariable: " << i << " gini=" << curr_gini << endl;
#endif

            if (curr_gini>maxgini)
            {
                maxgini=curr_gini;
                SplitVariable=i;
            }
        }

        // go over continuous variable
        for (int i=0; i<csplitDim; i++)
        {
            double curr_gini=continuousStatistics[i].ComputeGiniGain();
#ifdef DEBUG_PRINT

            cout << "\tVariable: " << (-i-1) << " gini=" << curr_gini << endl;
#endif

            if (curr_gini>maxgini)
            {
                maxgini=curr_gini;
                SplitVariable=-(i+1);
            }
        }

        if (maxgini==0.0)
            return false; // make the node a leaf, nobody can do a reasonable split

        // not set the split point for the variable picked as the split point
        if (SplitVariable>=0)
            splitSetProbability=discreteStatistics[SplitVariable].GetProbabilitySet();
        else
        {
            int splitVar=-SplitVariable-1;
            splitPoint=continuousStatistics[splitVar].GetSplit();
            splitSTD=sqrt(continuousStatistics[splitVar].getSplitVariance());
        }

#ifdef DEBUG_PRINT
        cout << "Split variable is " << SplitVariable << " and split point ";
        if (SplitVariable>=0)
            cout << splitSetProbability << endl;
        else
            cout << splitPoint << " " << splitSTD << endl;
#endif

        return true;
    }

    double ComputeProbabilityFirstClass(void)
    {
        return (mass0/mass);
    }

    bool MoreSplits(double minMass, int nodeId)
    {
        return ( mass>=minMass && mass0!=0.0 && mass0!=mass );
    }

    void SaveToStream(ostream& out)
    {
        out << SplitVariable << " ";
        if (SplitVariable>=0)
        {
            out << "( " << mass << " ) [ ";
            for (int i=0; i<splitSetProbability.size(); i++)
                out << splitSetProbability[i] << " ";
            out << " ] ";
        }
        else
        {
            out << "( " << splitPoint << " " << splitSTD << " ) ";
        }
    }
};

}

#endif //_CLUS_BINARYPROBABILISTICSPLITTER_H_

---