multidecisiontree.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_MULTIDECISIONTREE_H_
#define _CLUS_MULTIDECISIONTREE_H_

#include "machine.h"
#include "multidecisiontreenode.h"
#include "dctraingen.h"
#include "continuouslineartransformation.h"
#include "discretepermutationtransformation.h"
#include "statisticsgatherers.h"
#include "general.h"

#include <iostream>

// using namespace TNT;

namespace CLUS
{

template< class T_Splitter >
class MultiDecisionTree : public Machine
{
protected:
    MultiDecisionTreeNode< T_Splitter >* root;

   /// list of discrete domain sizes
    const Vector<int>& dDomainSize;

    /// num of discrete variables
    int dsplitDim;

    /// num of continuous+split variables
    int csplitDim;

    /// the number of trees built and outputs produces
    int noDatasets;
    
    /// the minimum number of datapoints in a node to split further
    int min_no_datapoints;

    /// type of split to be passed to splitter, splitter dependent
    int splitType;

    bool schemaMatch;
    bool labeled;
    bool alignSplits;

    ContinuousLinearTransformation* continuousTransformer;
    DiscretePermutationTransformation* discreteTransformer;
public:
    MultiDecisionTree(const Vector<int>& DDomainSize, int CsplitDim, int NoDatasets):
            Machine(CsplitDim,1),dDomainSize(DDomainSize),
            dsplitDim(DDomainSize.dim()-1),csplitDim(CsplitDim),noDatasets(NoDatasets)
    {
        ofstream file("x");
        continuousTransformer=0;
        discreteTransformer=0;
        min_no_datapoints = 10;
        splitType = 0;
    }

    ~MultiDecisionTree(void)
    {
        if (root!=0)
            delete root;

        if (continuousTransformer!=0)
            delete continuousTransformer;

        if (discreteTransformer!=0)
            delete discreteTransformer;
    }

    virtual int InDim(void)
    {
        return dsplitDim+csplitDim+1;
    }

    virtual string TypeName(void)
    {
        return string("MultiDecisionTree");
    }

    virtual void Infer(void)
    {
        if (root==0)
            return;
        // translate the first dsplitDim inputs into int
        int Dvars[MAX_VARIABLES];
        for(int i=0; i<dsplitDim; i++)
        {
            Dvars[i]=(int)(*input)[i];
            //    cout << Dvars[i] << " ";
        }
        //    cout << " :: ";
        double Cvars[MAX_VARIABLES];
        for (int i=0; i<csplitDim; i++)
        {
            Cvars[i]=(*input)[i+dsplitDim];
            //    cout << Cvars[i] << " ";
        }

        // shift the data
        int datasetIndex=(int)(*input)[dsplitDim+csplitDim];

        assert (datasetIndex<noDatasets);

        //    cout << " D:" << datasetIndex << " | " << endl;
        discreteTransformer->ApplyShiftToTuple(Dvars, datasetIndex);
        continuousTransformer->ApplyShiftToTuple(Cvars, datasetIndex);

        //     for(int i=0; i<dsplitDim; i++){
        //    cout << Dvars[i] << " ";
        //     }
        //    cout << " :: ";
        //   for (int i=0; i<csplitDim; i++){
        //cout << Cvars[i] << " ";
        //    }

        output[0]=root->Infer(Dvars, Cvars);
        //   cout << " O: " << output[0] << endl;
    }

    void printBitFlipData()
    {
        ofstream afile("hello");
        afile << "HELLO" << endl;
        bool significant[csplitDim];
        /*  for (int j=0; j<csplitDim; j++)
        {
        if (root->labeledMeansSignificant(j, 0))
        {
        significant[j]=true;
        }
        else significant[j] = false;
        }*/
        ofstream file("bitFlipStuff");
        bool bits[csplitDim];
        file << noDatasets;
        for (int j=0; j< csplitDim; j++)
        {
            bits[j]=root->negMeanLessThanPos(j, 0);
        }
        for (int i=1; i<noDatasets; i++)
        {
            file << "in bit flip loop" << endl;

            file << "Dataset " << i << ":" << endl;
            for (int j=0; j< csplitDim; j++)
            {
                if (root->labeledMeansSignificant(j,0))
                {
                    file << (root->negMeanLessThanPos(j, i)!= bits[j]) << endl;
                }
            }
        }
    }

    virtual void Identify()
    {
        // copy the data since we do not want to modify the original
        // The value of the last discrete attribute is the class label
        Matrix<double> ctrainData = training->GetTrainingData();

        continuousTransformer=new ContinuousLinearTransformation(csplitDim, noDatasets,
                              ctrainData) ;
        ofstream file("xyz");
        file << noDatasets << endl;
        // The value of the last continuous attribute designates the subdataset from
        // where the the tuple is comming from
        Matrix<int> dtrainData = dynamic_cast< DCTrainingData* >( training )
                                 -> GetDiscreteTrainingData();

        discreteTransformer=new DiscretePermutationTransformation(dsplitDim, noDatasets,
                            ctrainData, dtrainData,
                            dDomainSize);

        int M=ctrainData.num_rows();

        root = new MultiDecisionTreeNode<T_Splitter>
               (NULL, 1, dDomainSize, csplitDim, noDatasets,
                *discreteTransformer, *continuousTransformer);


        file << "here I am";
        file << noDatasets << endl;
        // ------------------- DOING SCHEMA MATCHING ---------------------
        if (schemaMatch)
        {
            //    cerr << "DOING SCHEMA MATCHING" << endl;
            // create the root and give it the Id 1

            root->StartLearningEpoch();
            for(int i=0; i<M; i++)
            {
                int datasetIndex=(int)ctrainData[i][csplitDim];

                ExitIf(datasetIndex>=noDatasets, "More datasets enountered than declared");
                int classLabel=dtrainData[i][dsplitDim];
                root->LearnSample(dtrainData[i],ctrainData[i], classLabel, datasetIndex);
            }
            file << noDatasets;
            //printBitFlipData(noDatasets);
            for (int att =0; att<dsplitDim; att++)
            { // discrete attribute

                // find and propagate the shift
                Vector<Permutation> shifts(noDatasets);
                shifts[0]=Permutation(dDomainSize[att]);
                for (int i=1; i<noDatasets; i++)
                    shifts[i]= root->ComputeDiscreteShift(true, att, i);
                discreteTransformer->SetShiftsAttribute(att, shifts);

            }

            for (int att =0; att<csplitDim; att++)
            { // continuous attribute

                double sumInvVars=0.0;
                double split0;
                if (labeled)
                {
                    split0 = 0.0;
                    //split0 = root->combineSplits(att, 0, &sumInvVars)/sumInvVars;
                }
                else
                {
                    if (!alignSplits)
                        split0= root->combineCenters(att, 0, &sumInvVars)/sumInvVars;
                    else
                        split0 == root->combineSplits(att, 0, &sumInvVars)/sumInvVars;
                }

                if (!NonZero(sumInvVars))
                    split0 =0.0;
                //        cerr << "split0";


                Vector<double> shifts(noDatasets);
                shifts[0]=0.0;
                for (int i=1; i<noDatasets; i++)
                {
                    double sumInvVars=0.0;
                    double spliti;
                    if (labeled)
                    {
                        // this aligns split points.  want to align centers
                        //spliti = root->combineSplits(att, i,&sumInvVars)/sumInvVars;
                        spliti = root->combineLabeledCenters(att, i, &sumInvVars)/sumInvVars;
                        if (!NonZero(sumInvVars))
                            spliti=split0;
                    }
                    else
                        spliti = root->combineCenters(att, i, &sumInvVars)/sumInvVars;
                    //          cerr << "here";

                    if (!NonZero(sumInvVars))
                        spliti=split0;
                    // else {
                    //        cerr << "sumInvVars=" << sumInvVars << endl;

                    shifts[i]= -(split0 - spliti);
                    //  }
                }
                continuousTransformer->SetShiftsAttribute(att, shifts);
            }

            // goto StartedLearning;
        }
        //      cerr << "finished schema matching" << endl;

        //  // create the root and give it the Id 1
        //root = new MultiDecisionTreeNode<T_Splitter>
        //    (1, dDomainSize, csplitDim, noDatasets,
        //     *discreteTransformer, *continuousTransformer);

        // learn in stages until nobody wants to learn anymore
        do
        {
            root->StartLearningEpoch();
            //printBitFlipData();
            for(int i=0; i<M; i++)
            {
                int datasetIndex=(int)ctrainData[i][csplitDim];

                ExitIf(datasetIndex>=noDatasets, "More datasets enountered than declared");

                int classLabel=dtrainData[i][dsplitDim];
                root->LearnSample(dtrainData[i],ctrainData[i], classLabel, datasetIndex);
            }
            //      StartedLearning:
            list<int> attList; // list of attributes; empty for now
            root->FindSplitAttributes(attList);
            // remove duplicates
            attList.sort();
            attList.unique();


            list<int>::iterator itrt;
            for (itrt=attList.begin();
                    itrt!=attList.end(); itrt++)
            {
                int SplitAttribute=*itrt;
                if (SplitAttribute>=0)
                { // discrete attribute

                    /*Vector< BinomialStatistics > statistics(noDatasets, BinomialStatistics(dDomainSize[SplitAttribute]) );
                      root->AddDiscreteShiftStatistics(SplitAttribute, statistics);
                      for (int i=0; i<noDatasets; i++)
                      statistics[i].CorrectWeightedStatistics(totalWeight);*/

                    // find and propagate the shift
                    Vector<Permutation> shifts(noDatasets);
                    shifts[0]=Permutation(dDomainSize[SplitAttribute]);
                    for (int i=1; i<noDatasets; i++)
                        shifts[i]= root->ComputeDiscreteShift(true, SplitAttribute, i);
                    discreteTransformer->SetShiftsAttribute(SplitAttribute, shifts);

                }
                else
                { // continuous attribute

                    /*Vector< NormalStatistics > statistics(noDatasets);
                      root->AddContinuousShiftStatistics(-SplitAttribute-1, statistics);
                      for (int i=0; i<noDatasets; i++)
                      statistics[i].CorrectWeightedStatistics(totalWeight);

                      // find and propagate the shift
                      Vector<double> shifts(noDatasets);
                      shifts[0]=0.0;
                      for (int i=1; i<noDatasets; i++)
                      shifts[i]=statistics[0].ComputeShift(labeled, statistics[i]);*/

                    double sumInvVars=0.0;
                    double split0;
                    if (labeled)
                    {
                        split0 = root->combineSplits(-SplitAttribute-1, 0, &sumInvVars)/sumInvVars;
                        split0=0.0;
                    }
                    else
                    {
                        if (!alignSplits)
                            split0 = root->combineCenters(-SplitAttribute-1, 0, &sumInvVars)/sumInvVars;
                        else
                            split0 = root->combineSplits(-SplitAttribute-1, 0, &sumInvVars)/sumInvVars;
                    }

                    if (sumInvVars==0.0)
                        split0=0.0;


                    // split0 = split0/ root->computeSumOfVarianceInverted(-SplitAttribute-1, 0);
                    Vector<double> shifts(noDatasets);
                    shifts[0]=0.0;
                    for (int i=1; i<noDatasets; i++)
                    {
                        double sumInvVars=0.0;
                        double spliti=0.0;
                        if (labeled)
                        {
                            // this aligns splitpnt.  want to align labeled centers
                            //spliti= root->combineSplits(-SplitAttribute-1, i,&sumInvVars)/sumInvVars;
                            shifts[i] = root->combineLabeledCenters(-SplitAttribute-1, i, &sumInvVars);
                        }
                        else
                        {
                            if (!alignSplits)
                                spliti= root->combineCenters(-SplitAttribute-1, i,&sumInvVars)/sumInvVars;
                            else
                                spliti = root->combineSplits(-SplitAttribute-1, i,&sumInvVars)/sumInvVars;

                        }
                        if (sumInvVars==0.0)
                            spliti=split0;
                        else
                        {
                            //        cerr << "sumInvVars=" << sumInvVars << endl;

                            //           cerr << "spliti " <<  spliti;
                            //           double denom = root->computeSumOfVarianceInverted(-SplitAttribute-1, i);
                            //           spliti = spliti/denom;
                            shifts[i]= -(split0 - spliti);
                        }
                    }
                    continuousTransformer->SetShiftsAttribute(-SplitAttribute-1, shifts);
                }
            }
        }
        while (root->StopLearningEpoch(splitType, min_no_datapoints));
        //      cout << "End Learning" << endl;
    }

    virtual void Prune(void)
    {
        const Matrix<double>& ctrainData = pruning->GetTrainingData();
        const Matrix<int>& dtrainData = dynamic_cast< DCTrainingData* >( pruning )
                                        -> GetDiscreteTrainingData();

        int M=ctrainData.num_rows();

        root->InitializePruningStatistics();
        for(int i=0; i<M; i++)
        {
            int datasetIndex=(int)ctrainData[i][csplitDim];
            // create copies of tuple
            int Dvars[MAX_VARIABLES];
            for(int j=0; j<dsplitDim; j++)
                Dvars[j]=dtrainData[i][j];

            double Cvars[MAX_VARIABLES];
            for (int j=0; j<csplitDim; j++)
                Cvars[j]=ctrainData[i][j];

            // apply the transformations directly to the data
            discreteTransformer->ApplyShiftToTuple(Dvars, datasetIndex);
            continuousTransformer->ApplyShiftToTuple(Cvars, datasetIndex);
            int classLabel=dtrainData[i][dsplitDim];
            root->UpdatePruningStatistics(Dvars, Cvars, classLabel, datasetIndex );
        }
        root->FinalizePruningStatistics();

        // now cut the tree to the right size
        double cost=root->PruneSubtree();
        cout << "RMSN after pruning is:" << cost/M << endl;
    }

    virtual int SetOption(char* name, char* val)
    {

        if (strcmp(name, "SchemaMatch")==0)
        {
            if (strcmp(val, "true")==0)
                schemaMatch=true;
            else
                schemaMatch=false;
        }
        else if (strcmp(name, "labeled")==0)
        {
            if (strcmp(val, "true")==0)
                labeled=true;
            else
                labeled=false;
        }
        else if (strcmp(name, "alignSplits")==0)
        {

            if (strcmp(val, "true")==0)
                alignSplits = true;
            else
                alignSplits=false;

        }
        else
            return Machine::SetOption(name,val);
        return 1;
    }

    virtual void SaveToStream(ostream& out)
    {
        // save the description in a stream
        out << "Tree: with "  << dsplitDim << " discrete and " << csplitDim << " continuous attributes" << endl;
        out << "Discrete attribute domain sizes: " << endl;

        out << "[ ";
        for(int i=0; i<dsplitDim; i++)
            out << dDomainSize[i] << " ";
        out << "]" << endl;

        out << "Shifts: " << endl;

        for(int i=0; i<dsplitDim; i++)
        {
            for (int j=1; j<noDatasets; j++)
            {
                out << "Shift to dataset " << j << " for discrete attribute " << i << ": " << endl;
                discreteTransformer->saveToStream(out, i, j);
            }
        }

        for(int i=0; i<csplitDim; i++)
        {
            for (int j=1; j < noDatasets; j++)
            {
                if (continuousTransformer->HasAttributeShifts(i))
                {
                    out << "Shift to dataset " << j << " for continuous attribute " << i << ": "
                    << continuousTransformer->getShift(i, j) << endl;
                }
                else
                {
                    out << "Shift to dataset " << j << " for continuous attribute " << i << ": NOSHIFT" << endl;
                }

            }
        }
        out << endl;
        root->SaveToStream(out);
    }

};
}

#endif // _CLUS_MULTIDECISIONTREE_H_

---