randomforest.cpp

   1 //
   2 //  randomforest.cpp
   3 //  Mothur
   4 //
   5 //  Created by Sarah Westcott on 10/2/12.
   6 //  Copyright (c) 2012 Schloss Lab. All rights reserved.
   7 //
   8
   9 #include "randomforest.hpp"
  10
  11 /***********************************************************************/
  12
  13 RandomForest::RandomForest(const vector <vector<int> > dataSet,
  14                            const int numDecisionTrees,
  15                            const string treeSplitCriterion = "gainratio",
  16                            const bool doPruning = false,
  17                            const float pruneAggressiveness = 0.9,
  18                            const bool discardHighErrorTrees = true,
  19                            const float highErrorTreeDiscardThreshold = 0.4,
  20                            const string optimumFeatureSubsetSelectionCriteria = "log2",
  21                            const float featureStandardDeviationThreshold = 0.0)
  22             : Forest(dataSet, numDecisionTrees, treeSplitCriterion, doPruning, pruneAggressiveness, discardHighErrorTrees, highErrorTreeDiscardThreshold, optimumFeatureSubsetSelectionCriteria, featureStandardDeviationThreshold) {
  23     m = MothurOut::getInstance();
  24 }
  25
  26 /***********************************************************************/
  27 // DONE
  28 int RandomForest::calcForrestErrorRate() {
  29     try {
  30         int numCorrect = 0;
  31         for (map<int, vector<int> >::iterator it = globalOutOfBagEstimates.begin(); it != globalOutOfBagEstimates.end(); it++) {
  32
  33             if (m->control_pressed) { return 0; }
  34
  35             int indexOfSample = it->first;
  36             vector<int> predictedOutComes = it->second;
  37             vector<int>::iterator maxPredictedOutComeIterator = max_element(predictedOutComes.begin(), predictedOutComes.end());
  38             int majorityVotedOutcome = (int)(maxPredictedOutComeIterator - predictedOutComes.begin());
  39             int realOutcome = dataSet[indexOfSample][numFeatures];
  40
  41             if (majorityVotedOutcome == realOutcome) { numCorrect++; }
  42         }
  43
  44         // TODO: save or return forrestErrorRate for future use;
  45         double forrestErrorRate = 1 - ((double)numCorrect / (double)globalOutOfBagEstimates.size());
  46
  47         m->mothurOut("numCorrect = " + toString(numCorrect)+ "\n");
  48         m->mothurOut("forrestErrorRate = " + toString(forrestErrorRate)+ "\n");
  49
  50         return 0;
  51     }
  52         catch(exception& e) {
  53                 m->errorOut(e, "RandomForest", "calcForrestErrorRate");
  54                 exit(1);
  55         }
  56 }
  57 /***********************************************************************/
  58
  59 int RandomForest::printConfusionMatrix(map<int, string> intToTreatmentMap) {
  60     try {
  61         int numGroups = intToTreatmentMap.size();
  62         vector<vector<int> > cm(numGroups, vector<int>(numGroups, 0));
  63
  64         for (map<int, vector<int> >::iterator it = globalOutOfBagEstimates.begin(); it != globalOutOfBagEstimates.end(); it++) {
  65
  66             if (m->control_pressed) { return 0; }
  67
  68             int indexOfSample = it->first; //key
  69             vector<int> predictedOutComes = it->second; //value, vector of all predicted classes
  70             vector<int>::iterator maxPredictedOutComeIterator = max_element(predictedOutComes.begin(), predictedOutComes.end());
  71             int majorityVotedOutcome = (int)(maxPredictedOutComeIterator - predictedOutComes.begin());
  72             int realOutcome = dataSet[indexOfSample][numFeatures];
  73             cm[realOutcome][majorityVotedOutcome] = cm[realOutcome][majorityVotedOutcome] + 1;
  74         }
  75
  76         vector<int> fw;
  77         for (int w = 0; w <numGroups; w++) {
  78             fw.push_back(intToTreatmentMap[w].length());
  79         }
  80
  81         m->mothurOut("confusion matrix:\n\t\t");
  82         for (int k = 0; k < numGroups; k++) {
  83             //m->mothurOut(intToTreatmentMap[k] + "\t");
  84             cout << setw(fw[k]) << intToTreatmentMap[k] << "\t";
  85         }
  86         for (int i = 0; i < numGroups; i++) {
  87             cout << "\n" << setw(fw[i]) << intToTreatmentMap[i] << "\t";
  88             //m->mothurOut("\n" + intToTreatmentMap[i] + "\t");
  89             if (m->control_pressed) { return 0; }
  90             for (int j = 0; j < numGroups; j++) {
  91                 //m->mothurOut(toString(cm[i][j]) + "\t");
  92                 cout << setw(fw[i]) << cm[i][j] << "\t";
  93             }
  94         }
  95         //m->mothurOut("\n");
  96         cout << "\n";
  97
  98         return 0;
  99     }
 100
 101     catch(exception& e) {
 102                 m->errorOut(e, "RandomForest", "printConfusionMatrix");
 103                 exit(1);
 104         }
 105 }
 106
 107 /***********************************************************************/
 108 int RandomForest::calcForrestVariableImportance(string filename) {
 109     try {
 110
 111         // follow the link: http://en.wikipedia.org/wiki/Dynamic_cast
 112         //if you are going to dynamically cast, aren't you undoing the advantage of abstraction. Why abstract at all?
 113         //could cause maintenance issues later if other types of Abstract decison trees are created that cannot be cast as a decision tree.
 114         for (int i = 0; i < decisionTrees.size(); i++) {
 115             if (m->control_pressed) { return 0; }
 116
 117             DecisionTree* decisionTree = dynamic_cast<DecisionTree*>(decisionTrees[i]);
 118
 119             for (int j = 0; j < numFeatures; j++) {
 120                 globalVariableImportanceList[j] += (double)decisionTree->variableImportanceList[j];
 121             }
 122         }
 123
 124         for (int i = 0;  i < numFeatures; i++) {
 125             globalVariableImportanceList[i] /= (double)numDecisionTrees;
 126         }
 127
 128         vector< pair<int, double> > globalVariableRanks;
 129         for (int i = 0; i < globalVariableImportanceList.size(); i++) {
 130             //cout << "[" << i << ',' << globalVariableImportanceList[i] << "], ";
 131             if (globalVariableImportanceList[i] > 0) {
 132                 pair<int, double> globalVariableRank(0, 0.0);
 133                 globalVariableRank.first = i;
 134                 globalVariableRank.second = globalVariableImportanceList[i];
 135                 globalVariableRanks.push_back(globalVariableRank);
 136             }
 137         }
 138
 139 //        for (int i = 0; i < globalVariableRanks.size(); i++) {
 140 //            cout << m->currentBinLabels[(int)globalVariableRanks[i][0]] << '\t' << globalVariableImportanceList[globalVariableRanks[i][0]] << endl;
 141 //        }
 142
 143
 144         VariableRankDescendingSorterDouble variableRankDescendingSorter;
 145         sort(globalVariableRanks.begin(), globalVariableRanks.end(), variableRankDescendingSorter);
 146
 147         ofstream out;
 148         m->openOutputFile(filename, out);
 149         out <<"OTU\tMean decrease accuracy\n";
 150         for (int i = 0; i < globalVariableRanks.size(); i++) {
 151             out << m->currentBinLabels[(int)globalVariableRanks[i].first] << '\t' << globalVariableImportanceList[globalVariableRanks[i].first] << endl;
 152         }
 153         out.close();
 154         return 0;
 155     }
 156         catch(exception& e) {
 157                 m->errorOut(e, "RandomForest", "calcForrestVariableImportance");
 158                 exit(1);
 159         }
 160 }
 161 /***********************************************************************/
 162 int RandomForest::populateDecisionTrees() {
 163     try {
 164
 165         vector<double> errorRateImprovements;
 166
 167         for (int i = 0; i < numDecisionTrees; i++) {
 168
 169             if (m->control_pressed) { return 0; }
 170             if (((i+1) % 100) == 0) {  m->mothurOut("Creating " + toString(i+1) + " (th) Decision tree\n");  }
 171
 172             // TODO: need to first fix if we are going to use pointer based system or anything else
 173             DecisionTree* decisionTree = new DecisionTree(dataSet, globalDiscardedFeatureIndices, OptimumFeatureSubsetSelector(optimumFeatureSubsetSelectionCriteria), treeSplitCriterion, featureStandardDeviationThreshold);
 174
 175             if (m->debug && doPruning) {
 176                 m->mothurOut("Before pruning\n");
 177                 decisionTree->printTree(decisionTree->rootNode, "ROOT");
 178             }
 179
 180             int numCorrect;
 181             double treeErrorRate;
 182
 183             decisionTree->calcTreeErrorRate(numCorrect, treeErrorRate);
 184             double prePrunedErrorRate = treeErrorRate;
 185
 186             if (m->debug) {
 187                 m->mothurOut("treeErrorRate: " + toString(treeErrorRate) + " numCorrect: " + toString(numCorrect) + "\n");
 188             }
 189
 190             if (doPruning) {
 191                 decisionTree->pruneTree(pruneAggressiveness);
 192                 if (m->debug) {
 193                     m->mothurOut("After pruning\n");
 194                     decisionTree->printTree(decisionTree->rootNode, "ROOT");
 195                 }
 196                 decisionTree->calcTreeErrorRate(numCorrect, treeErrorRate);
 197             }
 198             double postPrunedErrorRate = treeErrorRate;
 199
 200
 201             decisionTree->calcTreeVariableImportanceAndError(numCorrect, treeErrorRate);
 202             double errorRateImprovement = (prePrunedErrorRate - postPrunedErrorRate) / prePrunedErrorRate;
 203
 204             if (m->debug) {
 205                 m->mothurOut("treeErrorRate: " + toString(treeErrorRate) + " numCorrect: " + toString(numCorrect) + "\n");
 206                 if (doPruning) {
 207                     m->mothurOut("errorRateImprovement: " + toString(errorRateImprovement) + "\n");
 208                 }
 209             }
 210
 211
 212             if (discardHighErrorTrees) {
 213                 if (treeErrorRate < highErrorTreeDiscardThreshold) {
 214                     updateGlobalOutOfBagEstimates(decisionTree);
 215                     decisionTree->purgeDataSetsFromTree();
 216                     decisionTrees.push_back(decisionTree);
 217                     if (doPruning) {
 218                         errorRateImprovements.push_back(errorRateImprovement);
 219                     }
 220                 } else {
 221                     delete decisionTree;
 222                 }
 223             } else {
 224                 updateGlobalOutOfBagEstimates(decisionTree);
 225                 decisionTree->purgeDataSetsFromTree();
 226                 decisionTrees.push_back(decisionTree);
 227                 if (doPruning) {
 228                     errorRateImprovements.push_back(errorRateImprovement);
 229                 }
 230             }
 231         }
 232
 233         double avgErrorRateImprovement = -1.0;
 234         if (errorRateImprovements.size() > 0) {
 235             avgErrorRateImprovement = accumulate(errorRateImprovements.begin(), errorRateImprovements.end(), 0.0);
 236 //            cout << "Total " << avgErrorRateImprovement << " size " << errorRateImprovements.size() << endl;
 237             avgErrorRateImprovement /= errorRateImprovements.size();
 238         }
 239
 240         if (m->debug && doPruning) {
 241             m->mothurOut("avgErrorRateImprovement:" + toString(avgErrorRateImprovement) + "\n");
 242         }
 243         // m->mothurOut("globalOutOfBagEstimates = " + toStringVectorMap(globalOutOfBagEstimates)+ "\n");
 244
 245
 246         return 0;
 247     }
 248     catch(exception& e) {
 249         m->errorOut(e, "RandomForest", "populateDecisionTrees");
 250         exit(1);
 251     }
 252 }
 253 /***********************************************************************/
 254 // TODO: need to finalize bettween reference and pointer for DecisionTree [partially solved]
 255 // DONE: make this pure virtual in superclass
 256 // DONE
 257 int RandomForest::updateGlobalOutOfBagEstimates(DecisionTree* decisionTree) {
 258     try {
 259         for (map<int, int>::iterator it = decisionTree->outOfBagEstimates.begin(); it != decisionTree->outOfBagEstimates.end(); it++) {
 260
 261             if (m->control_pressed) { return 0; }
 262
 263             int indexOfSample = it->first;
 264             int predictedOutcomeOfSample = it->second;
 265
 266             if (globalOutOfBagEstimates.count(indexOfSample) == 0) {
 267                 globalOutOfBagEstimates[indexOfSample] = vector<int>(decisionTree->numOutputClasses, 0);
 268             };
 269
 270             globalOutOfBagEstimates[indexOfSample][predictedOutcomeOfSample] += 1;
 271         }
 272         return 0;
 273     }
 274     catch(exception& e) {
 275         m->errorOut(e, "RandomForest", "updateGlobalOutOfBagEstimates");
 276         exit(1);
 277     }
 278 }
 279 /***********************************************************************/
 280
 281