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         m->mothurOut("confusion matrix:\n\t\t");
  77         for (int i = 0; i < numGroups; i++) {
  78             m->mothurOut(intToTreatmentMap[i] + "\t");
  79         }
  80         for (int i = 0; i < numGroups; i++) {
  81             //m->mothurOut("\n" + intToTreatmentMap[i] + "\t");
  82             if (m->control_pressed) { return 0; }
  83             for (int j = 0; j < numGroups; j++) {
  84                 m->mothurOut(cm[i][j] + "\t");
  85             }
  86         }
  87         m->mothurOut("\n");
  88
  89         return 0;
  90     }
  91
  92     catch(exception& e) {
  93                 m->errorOut(e, "RandomForest", "printConfusionMatrix");
  94                 exit(1);
  95         }
  96 }
  97
  98 /***********************************************************************/
  99 int RandomForest::calcForrestVariableImportance(string filename) {
 100     try {
 101
 102         // follow the link: http://en.wikipedia.org/wiki/Dynamic_cast
 103         //if you are going to dynamically cast, aren't you undoing the advantage of abstraction. Why abstract at all?
 104         //could cause maintenance issues later if other types of Abstract decison trees are created that cannot be cast as a decision tree.
 105         for (int i = 0; i < decisionTrees.size(); i++) {
 106             if (m->control_pressed) { return 0; }
 107
 108             DecisionTree* decisionTree = dynamic_cast<DecisionTree*>(decisionTrees[i]);
 109
 110             for (int j = 0; j < numFeatures; j++) {
 111                 globalVariableImportanceList[j] += (double)decisionTree->variableImportanceList[j];
 112             }
 113         }
 114
 115         for (int i = 0;  i < numFeatures; i++) {
 116             globalVariableImportanceList[i] /= (double)numDecisionTrees;
 117         }
 118
 119         vector< pair<int, double> > globalVariableRanks;
 120         for (int i = 0; i < globalVariableImportanceList.size(); i++) {
 121             //cout << "[" << i << ',' << globalVariableImportanceList[i] << "], ";
 122             if (globalVariableImportanceList[i] > 0) {
 123                 pair<int, double> globalVariableRank(0, 0.0);
 124                 globalVariableRank.first = i;
 125                 globalVariableRank.second = globalVariableImportanceList[i];
 126                 globalVariableRanks.push_back(globalVariableRank);
 127             }
 128         }
 129
 130 //        for (int i = 0; i < globalVariableRanks.size(); i++) {
 131 //            cout << m->currentBinLabels[(int)globalVariableRanks[i][0]] << '\t' << globalVariableImportanceList[globalVariableRanks[i][0]] << endl;
 132 //        }
 133
 134
 135         VariableRankDescendingSorterDouble variableRankDescendingSorter;
 136         sort(globalVariableRanks.begin(), globalVariableRanks.end(), variableRankDescendingSorter);
 137
 138         ofstream out;
 139         m->openOutputFile(filename, out);
 140         out <<"OTU\tMean decrease accuracy\n";
 141         for (int i = 0; i < globalVariableRanks.size(); i++) {
 142             out << m->currentBinLabels[(int)globalVariableRanks[i].first] << '\t' << globalVariableImportanceList[globalVariableRanks[i].first] << endl;
 143         }
 144         out.close();
 145         return 0;
 146     }
 147         catch(exception& e) {
 148                 m->errorOut(e, "RandomForest", "calcForrestVariableImportance");
 149                 exit(1);
 150         }
 151 }
 152 /***********************************************************************/
 153 int RandomForest::populateDecisionTrees() {
 154     try {
 155
 156         vector<double> errorRateImprovements;
 157
 158         for (int i = 0; i < numDecisionTrees; i++) {
 159
 160             if (m->control_pressed) { return 0; }
 161             if (((i+1) % 100) == 0) {  m->mothurOut("Creating " + toString(i+1) + " (th) Decision tree\n");  }
 162
 163             // TODO: need to first fix if we are going to use pointer based system or anything else
 164             DecisionTree* decisionTree = new DecisionTree(dataSet, globalDiscardedFeatureIndices, OptimumFeatureSubsetSelector(optimumFeatureSubsetSelectionCriteria), treeSplitCriterion, featureStandardDeviationThreshold);
 165
 166             if (m->debug && doPruning) {
 167                 m->mothurOut("Before pruning\n");
 168                 decisionTree->printTree(decisionTree->rootNode, "ROOT");
 169             }
 170
 171             int numCorrect;
 172             double treeErrorRate;
 173
 174             decisionTree->calcTreeErrorRate(numCorrect, treeErrorRate);
 175             double prePrunedErrorRate = treeErrorRate;
 176
 177             if (m->debug) {
 178                 m->mothurOut("treeErrorRate: " + toString(treeErrorRate) + " numCorrect: " + toString(numCorrect) + "\n");
 179             }
 180
 181             if (doPruning) {
 182                 decisionTree->pruneTree(pruneAggressiveness);
 183                 if (m->debug) {
 184                     m->mothurOut("After pruning\n");
 185                     decisionTree->printTree(decisionTree->rootNode, "ROOT");
 186                 }
 187                 decisionTree->calcTreeErrorRate(numCorrect, treeErrorRate);
 188             }
 189             double postPrunedErrorRate = treeErrorRate;
 190
 191
 192             decisionTree->calcTreeVariableImportanceAndError(numCorrect, treeErrorRate);
 193             double errorRateImprovement = (prePrunedErrorRate - postPrunedErrorRate) / prePrunedErrorRate;
 194
 195             if (m->debug) {
 196                 m->mothurOut("treeErrorRate: " + toString(treeErrorRate) + " numCorrect: " + toString(numCorrect) + "\n");
 197                 if (doPruning) {
 198                     m->mothurOut("errorRateImprovement: " + toString(errorRateImprovement) + "\n");
 199                 }
 200             }
 201
 202
 203             if (discardHighErrorTrees) {
 204                 if (treeErrorRate < highErrorTreeDiscardThreshold) {
 205                     updateGlobalOutOfBagEstimates(decisionTree);
 206                     decisionTree->purgeDataSetsFromTree();
 207                     decisionTrees.push_back(decisionTree);
 208                     if (doPruning) {
 209                         errorRateImprovements.push_back(errorRateImprovement);
 210                     }
 211                 } else {
 212                     delete decisionTree;
 213                 }
 214             } else {
 215                 updateGlobalOutOfBagEstimates(decisionTree);
 216                 decisionTree->purgeDataSetsFromTree();
 217                 decisionTrees.push_back(decisionTree);
 218                 if (doPruning) {
 219                     errorRateImprovements.push_back(errorRateImprovement);
 220                 }
 221             }
 222         }
 223
 224         double avgErrorRateImprovement = -1.0;
 225         if (errorRateImprovements.size() > 0) {
 226             avgErrorRateImprovement = accumulate(errorRateImprovements.begin(), errorRateImprovements.end(), 0.0);
 227 //            cout << "Total " << avgErrorRateImprovement << " size " << errorRateImprovements.size() << endl;
 228             avgErrorRateImprovement /= errorRateImprovements.size();
 229         }
 230
 231         if (m->debug && doPruning) {
 232             m->mothurOut("avgErrorRateImprovement:" + toString(avgErrorRateImprovement) + "\n");
 233         }
 234         // m->mothurOut("globalOutOfBagEstimates = " + toStringVectorMap(globalOutOfBagEstimates)+ "\n");
 235
 236
 237         return 0;
 238     }
 239     catch(exception& e) {
 240         m->errorOut(e, "RandomForest", "populateDecisionTrees");
 241         exit(1);
 242     }
 243 }
 244 /***********************************************************************/
 245 // TODO: need to finalize bettween reference and pointer for DecisionTree [partially solved]
 246 // DONE: make this pure virtual in superclass
 247 // DONE
 248 int RandomForest::updateGlobalOutOfBagEstimates(DecisionTree* decisionTree) {
 249     try {
 250         for (map<int, int>::iterator it = decisionTree->outOfBagEstimates.begin(); it != decisionTree->outOfBagEstimates.end(); it++) {
 251
 252             if (m->control_pressed) { return 0; }
 253
 254             int indexOfSample = it->first;
 255             int predictedOutcomeOfSample = it->second;
 256
 257             if (globalOutOfBagEstimates.count(indexOfSample) == 0) {
 258                 globalOutOfBagEstimates[indexOfSample] = vector<int>(decisionTree->numOutputClasses, 0);
 259             };
 260
 261             globalOutOfBagEstimates[indexOfSample][predictedOutcomeOfSample] += 1;
 262         }
 263         return 0;
 264     }
 265     catch(exception& e) {
 266         m->errorOut(e, "RandomForest", "updateGlobalOutOfBagEstimates");
 267         exit(1);
 268     }
 269 }
 270 /***********************************************************************/
 271
 272