decision_tree.cc

   1
   2 ///////////////////////////////////////////////////////////////////////////
   3 // This program is free software: you can redistribute it and/or modify  //
   4 // it under the terms of the version 3 of the GNU General Public License //
   5 // as published by the Free Software Foundation.                         //
   6 //                                                                       //
   7 // This program is distributed in the hope that it will be useful, but   //
   8 // WITHOUT ANY WARRANTY; without even the implied warranty of            //
   9 // MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU      //
  10 // General Public License for more details.                              //
  11 //                                                                       //
  12 // You should have received a copy of the GNU General Public License     //
  13 // along with this program. If not, see <http://www.gnu.org/licenses/>.  //
  14 //                                                                       //
  15 // Written by Francois Fleuret, (C) IDIAP                                //
  16 // Contact <francois.fleuret@idiap.ch> for comments & bug reports        //
  17 ///////////////////////////////////////////////////////////////////////////
  18
  19 #include "decision_tree.h"
  20 #include "fusion_sort.h"
  21
  22 DecisionTree::DecisionTree() {
  23   _feature_index = -1;
  24   _threshold = 0;
  25   _weight = 0;
  26   _subtree_greater = 0;
  27   _subtree_lesser = 0;
  28 }
  29
  30 DecisionTree::~DecisionTree() {
  31   if(_subtree_lesser)
  32     delete _subtree_lesser;
  33   if(_subtree_greater)
  34     delete _subtree_greater;
  35 }
  36
  37 int DecisionTree::nb_leaves() {
  38   if(_subtree_lesser ||_subtree_greater)
  39     return _subtree_lesser->nb_leaves() + _subtree_greater->nb_leaves();
  40   else
  41     return 1;
  42 }
  43
  44 int DecisionTree::depth() {
  45   if(_subtree_lesser ||_subtree_greater)
  46     return 1 + max(_subtree_lesser->depth(), _subtree_greater->depth());
  47   else
  48     return 1;
  49 }
  50
  51 scalar_t DecisionTree::response(SampleSet *sample_set, int n_sample) {
  52   if(_subtree_lesser && _subtree_greater) {
  53     if(sample_set->feature_value(n_sample, _feature_index) < _threshold)
  54       return _subtree_lesser->response(sample_set, n_sample);
  55     else
  56       return _subtree_greater->response(sample_set, n_sample);
  57   } else {
  58     return _weight;
  59   }
  60 }
  61
  62 void DecisionTree::pick_best_split(SampleSet *sample_set, scalar_t *loss_derivatives) {
  63
  64   int nb_samples = sample_set->nb_samples();
  65
  66   scalar_t *responses = new scalar_t[nb_samples];
  67   int *indexes = new int[nb_samples];
  68   int *sorted_indexes = new int[nb_samples];
  69
  70   scalar_t max_abs_sum = 0;
  71   _feature_index = -1;
  72
  73   for(int f = 0; f < sample_set->nb_features(); f++) {
  74     scalar_t sum = 0;
  75
  76     for(int s = 0; s < nb_samples; s++) {
  77       indexes[s] = s;
  78       responses[s] = sample_set->feature_value(s, f);
  79       sum += loss_derivatives[s];
  80     }
  81
  82     indexed_fusion_sort(nb_samples, indexes, sorted_indexes, responses);
  83
  84     int t, u = sorted_indexes[0];
  85     for(int s = 0; s < nb_samples - 1; s++) {
  86       t = u;
  87       u = sorted_indexes[s + 1];
  88       sum -= 2 * loss_derivatives[t];
  89
  90       if(responses[t] < responses[u] && abs(sum) > max_abs_sum) {
  91         max_abs_sum = abs(sum);
  92         _feature_index = f;
  93         _threshold = (responses[t] + responses[u])/2;
  94       }
  95     }
  96   }
  97
  98   delete[] indexes;
  99   delete[] sorted_indexes;
 100   delete[] responses;
 101 }
 102
 103 void DecisionTree::train(LossMachine *loss_machine,
 104                          SampleSet *sample_set,
 105                          scalar_t *current_responses,
 106                          scalar_t *loss_derivatives,
 107                          int depth) {
 108
 109   if(_subtree_lesser || _subtree_greater || _feature_index >= 0) {
 110     cerr << "You can not re-train a tree." << endl;
 111     abort();
 112   }
 113
 114   int nb_samples = sample_set->nb_samples();
 115
 116   int nb_pos = 0, nb_neg = 0;
 117   for(int s = 0; s < sample_set->nb_samples(); s++) {
 118     if(sample_set->label(s) > 0) nb_pos++;
 119     else if(sample_set->label(s) < 0) nb_neg++;
 120   }
 121
 122   (*global.log_stream) << "Training tree" << endl;
 123   (*global.log_stream) << "  nb_samples " << nb_samples << endl;
 124   (*global.log_stream) << "  depth " << depth << endl;
 125   (*global.log_stream) << "  nb_pos = " << nb_pos << endl;
 126   (*global.log_stream) << "  nb_neg = " << nb_neg << endl;
 127
 128   if(depth >= global.tree_depth_max)
 129     (*global.log_stream) << "  Maximum depth reached." << endl;
 130   if(nb_pos < min_nb_samples_for_split)
 131     (*global.log_stream) << "  Not enough positive samples." << endl;
 132   if(nb_neg < min_nb_samples_for_split)
 133     (*global.log_stream) << "  Not enough negative samples." << endl;
 134
 135   if(depth < global.tree_depth_max &&
 136      nb_pos >= min_nb_samples_for_split &&
 137      nb_neg >= min_nb_samples_for_split) {
 138
 139     pick_best_split(sample_set, loss_derivatives);
 140
 141     if(_feature_index >= 0) {
 142       int indexes[nb_samples];
 143       scalar_t *parted_current_responses = new scalar_t[nb_samples];
 144       scalar_t *parted_loss_derivatives = new scalar_t[nb_samples];
 145
 146       int nb_lesser = 0, nb_greater = 0;
 147       int nb_lesser_pos = 0, nb_lesser_neg = 0, nb_greater_pos = 0, nb_greater_neg = 0;
 148
 149       for(int s = 0; s < nb_samples; s++) {
 150         if(sample_set->feature_value(s, _feature_index) < _threshold) {
 151           indexes[nb_lesser] = s;
 152           parted_current_responses[nb_lesser] = current_responses[s];
 153           parted_loss_derivatives[nb_lesser] = loss_derivatives[s];
 154
 155           if(sample_set->label(s) > 0)
 156             nb_lesser_pos++;
 157           else if(sample_set->label(s) < 0)
 158             nb_lesser_neg++;
 159
 160           nb_lesser++;
 161         } else {
 162           nb_greater++;
 163
 164           indexes[nb_samples - nb_greater] = s;
 165           parted_current_responses[nb_samples - nb_greater] = current_responses[s];
 166           parted_loss_derivatives[nb_samples - nb_greater] = loss_derivatives[s];
 167
 168           if(sample_set->label(s) > 0)
 169             nb_greater_pos++;
 170           else if(sample_set->label(s) < 0)
 171             nb_greater_neg++;
 172         }
 173       }
 174
 175       if((nb_lesser_pos >= min_nb_samples_for_split ||
 176           nb_lesser_neg >= min_nb_samples_for_split) &&
 177          (nb_greater_pos >= min_nb_samples_for_split ||
 178           nb_greater_neg >= min_nb_samples_for_split)) {
 179
 180         _subtree_lesser = new DecisionTree();
 181
 182         {
 183           SampleSet sub_sample_set(sample_set, nb_lesser, indexes);
 184
 185           _subtree_lesser->train(loss_machine,
 186                                  &sub_sample_set,
 187                                  parted_current_responses,
 188                                  parted_loss_derivatives,
 189                                  depth + 1);
 190         }
 191
 192         _subtree_greater = new DecisionTree();
 193
 194         {
 195           SampleSet sub_sample_set(sample_set, nb_greater, indexes + nb_lesser);
 196
 197           _subtree_greater->train(loss_machine,
 198                                   &sub_sample_set,
 199                                   parted_current_responses + nb_lesser,
 200                                   parted_loss_derivatives + nb_lesser,
 201                                   depth + 1);
 202         }
 203       }
 204
 205       delete[] parted_current_responses;
 206       delete[] parted_loss_derivatives;
 207     } else {
 208       (*global.log_stream) << "Could not find a feature for split." << endl;
 209     }
 210   }
 211
 212   if(!(_subtree_greater && _subtree_lesser)) {
 213     scalar_t *tmp_responses = new scalar_t[nb_samples];
 214     for(int s = 0; s < nb_samples; s++)
 215       tmp_responses[s] = 1;
 216
 217     _weight = loss_machine->optimal_weight(sample_set, tmp_responses, current_responses);
 218
 219     const scalar_t max_weight = 10.0;
 220
 221     if(_weight > max_weight) {
 222       _weight = max_weight;
 223     } else if(_weight < - max_weight) {
 224       _weight = - max_weight;
 225     }
 226
 227     (*global.log_stream) << "  _weight " << _weight << endl;
 228
 229     delete[] tmp_responses;
 230   }
 231 }
 232
 233 void DecisionTree::train(LossMachine *loss_machine,
 234                  SampleSet *sample_set,
 235                  scalar_t *current_responses) {
 236
 237   scalar_t *loss_derivatives = new scalar_t[sample_set->nb_samples()];
 238
 239   loss_machine->get_loss_derivatives(sample_set, current_responses, loss_derivatives);
 240
 241   train(loss_machine, sample_set, current_responses, loss_derivatives, 0);
 242
 243   delete[] loss_derivatives;
 244 }
 245
 246 //////////////////////////////////////////////////////////////////////
 247
 248 void DecisionTree::tag_used_features(bool *used) {
 249   if(_subtree_lesser && _subtree_greater) {
 250     used[_feature_index] = true;
 251     _subtree_lesser->tag_used_features(used);
 252     _subtree_greater->tag_used_features(used);
 253   }
 254 }
 255
 256 void DecisionTree::re_index_features(int *new_indexes) {
 257   if(_subtree_lesser && _subtree_greater) {
 258     _feature_index = new_indexes[_feature_index];
 259     _subtree_lesser->re_index_features(new_indexes);
 260     _subtree_greater->re_index_features(new_indexes);
 261   }
 262 }
 263
 264 //////////////////////////////////////////////////////////////////////
 265
 266 void DecisionTree::read(istream *is) {
 267   if(_subtree_lesser || _subtree_greater) {
 268     cerr << "You can not read in an existing tree." << endl;
 269     abort();
 270   }
 271
 272   read_var(is, &_feature_index);
 273   read_var(is, &_threshold);
 274   read_var(is, &_weight);
 275
 276   int split;
 277   read_var(is, &split);
 278
 279   if(split) {
 280     _subtree_lesser = new DecisionTree();
 281     _subtree_lesser->read(is);
 282     _subtree_greater = new DecisionTree();
 283     _subtree_greater->read(is);
 284   }
 285 }
 286
 287 void DecisionTree::write(ostream *os) {
 288
 289   write_var(os, &_feature_index);
 290   write_var(os, &_threshold);
 291   write_var(os, &_weight);
 292
 293   int split;
 294   if(_subtree_lesser && _subtree_greater) {
 295     split = 1;
 296     write_var(os, &split);
 297     _subtree_lesser->write(os);
 298     _subtree_greater->write(os);
 299   } else {
 300     split = 0;
 301     write_var(os, &split);
 302   }
 303 }