[29] | 1 | /* |
---|
| 2 | * This program is free software; you can redistribute it and/or modify |
---|
| 3 | * it under the terms of the GNU General Public License as published by |
---|
| 4 | * the Free Software Foundation; either version 2 of the License, or |
---|
| 5 | * (at your option) any later version. |
---|
| 6 | * |
---|
| 7 | * This program is distributed in the hope that it will be useful, |
---|
| 8 | * but WITHOUT ANY WARRANTY; without even the implied warranty of |
---|
| 9 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
---|
| 10 | * GNU 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, write to the Free Software |
---|
| 14 | * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. |
---|
| 15 | */ |
---|
| 16 | |
---|
| 17 | /* |
---|
| 18 | * LogitBoost.java |
---|
| 19 | * Copyright (C) 1999, 2002 University of Waikato, Hamilton, New Zealand |
---|
| 20 | * |
---|
| 21 | */ |
---|
| 22 | |
---|
| 23 | package weka.classifiers.meta; |
---|
| 24 | |
---|
| 25 | import weka.classifiers.Classifier; |
---|
| 26 | import weka.classifiers.AbstractClassifier; |
---|
| 27 | import weka.classifiers.Evaluation; |
---|
| 28 | import weka.classifiers.RandomizableIteratedSingleClassifierEnhancer; |
---|
| 29 | import weka.classifiers.Sourcable; |
---|
| 30 | import weka.core.Attribute; |
---|
| 31 | import weka.core.Capabilities; |
---|
| 32 | import weka.core.Instance; |
---|
| 33 | import weka.core.Instances; |
---|
| 34 | import weka.core.Option; |
---|
| 35 | import weka.core.RevisionUtils; |
---|
| 36 | import weka.core.TechnicalInformation; |
---|
| 37 | import weka.core.TechnicalInformationHandler; |
---|
| 38 | import weka.core.Utils; |
---|
| 39 | import weka.core.WeightedInstancesHandler; |
---|
| 40 | import weka.core.Capabilities.Capability; |
---|
| 41 | import weka.core.TechnicalInformation.Field; |
---|
| 42 | import weka.core.TechnicalInformation.Type; |
---|
| 43 | |
---|
| 44 | import java.util.Enumeration; |
---|
| 45 | import java.util.Random; |
---|
| 46 | import java.util.Vector; |
---|
| 47 | |
---|
| 48 | /** |
---|
| 49 | <!-- globalinfo-start --> |
---|
| 50 | * Class for performing additive logistic regression. <br/> |
---|
| 51 | * This class performs classification using a regression scheme as the base learner, and can handle multi-class problems. For more information, see<br/> |
---|
| 52 | * <br/> |
---|
| 53 | * J. Friedman, T. Hastie, R. Tibshirani (1998). Additive Logistic Regression: a Statistical View of Boosting. Stanford University.<br/> |
---|
| 54 | * <br/> |
---|
| 55 | * Can do efficient internal cross-validation to determine appropriate number of iterations. |
---|
| 56 | * <p/> |
---|
| 57 | <!-- globalinfo-end --> |
---|
| 58 | * |
---|
| 59 | <!-- technical-bibtex-start --> |
---|
| 60 | * BibTeX: |
---|
| 61 | * <pre> |
---|
| 62 | * @techreport{Friedman1998, |
---|
| 63 | * address = {Stanford University}, |
---|
| 64 | * author = {J. Friedman and T. Hastie and R. Tibshirani}, |
---|
| 65 | * title = {Additive Logistic Regression: a Statistical View of Boosting}, |
---|
| 66 | * year = {1998}, |
---|
| 67 | * PS = {http://www-stat.stanford.edu/\~jhf/ftp/boost.ps} |
---|
| 68 | * } |
---|
| 69 | * </pre> |
---|
| 70 | * <p/> |
---|
| 71 | <!-- technical-bibtex-end --> |
---|
| 72 | * |
---|
| 73 | <!-- options-start --> |
---|
| 74 | * Valid options are: <p/> |
---|
| 75 | * |
---|
| 76 | * <pre> -Q |
---|
| 77 | * Use resampling instead of reweighting for boosting.</pre> |
---|
| 78 | * |
---|
| 79 | * <pre> -P <percent> |
---|
| 80 | * Percentage of weight mass to base training on. |
---|
| 81 | * (default 100, reduce to around 90 speed up)</pre> |
---|
| 82 | * |
---|
| 83 | * <pre> -F <num> |
---|
| 84 | * Number of folds for internal cross-validation. |
---|
| 85 | * (default 0 -- no cross-validation)</pre> |
---|
| 86 | * |
---|
| 87 | * <pre> -R <num> |
---|
| 88 | * Number of runs for internal cross-validation. |
---|
| 89 | * (default 1)</pre> |
---|
| 90 | * |
---|
| 91 | * <pre> -L <num> |
---|
| 92 | * Threshold on the improvement of the likelihood. |
---|
| 93 | * (default -Double.MAX_VALUE)</pre> |
---|
| 94 | * |
---|
| 95 | * <pre> -H <num> |
---|
| 96 | * Shrinkage parameter. |
---|
| 97 | * (default 1)</pre> |
---|
| 98 | * |
---|
| 99 | * <pre> -S <num> |
---|
| 100 | * Random number seed. |
---|
| 101 | * (default 1)</pre> |
---|
| 102 | * |
---|
| 103 | * <pre> -I <num> |
---|
| 104 | * Number of iterations. |
---|
| 105 | * (default 10)</pre> |
---|
| 106 | * |
---|
| 107 | * <pre> -D |
---|
| 108 | * If set, classifier is run in debug mode and |
---|
| 109 | * may output additional info to the console</pre> |
---|
| 110 | * |
---|
| 111 | * <pre> -W |
---|
| 112 | * Full name of base classifier. |
---|
| 113 | * (default: weka.classifiers.trees.DecisionStump)</pre> |
---|
| 114 | * |
---|
| 115 | * <pre> |
---|
| 116 | * Options specific to classifier weka.classifiers.trees.DecisionStump: |
---|
| 117 | * </pre> |
---|
| 118 | * |
---|
| 119 | * <pre> -D |
---|
| 120 | * If set, classifier is run in debug mode and |
---|
| 121 | * may output additional info to the console</pre> |
---|
| 122 | * |
---|
| 123 | <!-- options-end --> |
---|
| 124 | * |
---|
| 125 | * Options after -- are passed to the designated learner.<p> |
---|
| 126 | * |
---|
| 127 | * @author Len Trigg (trigg@cs.waikato.ac.nz) |
---|
| 128 | * @author Eibe Frank (eibe@cs.waikato.ac.nz) |
---|
| 129 | * @version $Revision: 6091 $ |
---|
| 130 | */ |
---|
| 131 | public class LogitBoost |
---|
| 132 | extends RandomizableIteratedSingleClassifierEnhancer |
---|
| 133 | implements Sourcable, WeightedInstancesHandler, TechnicalInformationHandler { |
---|
| 134 | |
---|
| 135 | /** for serialization */ |
---|
| 136 | static final long serialVersionUID = -3905660358715833753L; |
---|
| 137 | |
---|
| 138 | /** Array for storing the generated base classifiers. |
---|
| 139 | Note: we are hiding the variable from IteratedSingleClassifierEnhancer*/ |
---|
| 140 | protected Classifier [][] m_Classifiers; |
---|
| 141 | |
---|
| 142 | /** The number of classes */ |
---|
| 143 | protected int m_NumClasses; |
---|
| 144 | |
---|
| 145 | /** The number of successfully generated base classifiers. */ |
---|
| 146 | protected int m_NumGenerated; |
---|
| 147 | |
---|
| 148 | /** The number of folds for the internal cross-validation. */ |
---|
| 149 | protected int m_NumFolds = 0; |
---|
| 150 | |
---|
| 151 | /** The number of runs for the internal cross-validation. */ |
---|
| 152 | protected int m_NumRuns = 1; |
---|
| 153 | |
---|
| 154 | /** Weight thresholding. The percentage of weight mass used in training */ |
---|
| 155 | protected int m_WeightThreshold = 100; |
---|
| 156 | |
---|
| 157 | /** A threshold for responses (Friedman suggests between 2 and 4) */ |
---|
| 158 | protected static final double Z_MAX = 3; |
---|
| 159 | |
---|
| 160 | /** Dummy dataset with a numeric class */ |
---|
| 161 | protected Instances m_NumericClassData; |
---|
| 162 | |
---|
| 163 | /** The actual class attribute (for getting class names) */ |
---|
| 164 | protected Attribute m_ClassAttribute; |
---|
| 165 | |
---|
| 166 | /** Use boosting with reweighting? */ |
---|
| 167 | protected boolean m_UseResampling; |
---|
| 168 | |
---|
| 169 | /** The threshold on the improvement of the likelihood */ |
---|
| 170 | protected double m_Precision = -Double.MAX_VALUE; |
---|
| 171 | |
---|
| 172 | /** The value of the shrinkage parameter */ |
---|
| 173 | protected double m_Shrinkage = 1; |
---|
| 174 | |
---|
| 175 | /** The random number generator used */ |
---|
| 176 | protected Random m_RandomInstance = null; |
---|
| 177 | |
---|
| 178 | /** The value by which the actual target value for the |
---|
| 179 | true class is offset. */ |
---|
| 180 | protected double m_Offset = 0.0; |
---|
| 181 | |
---|
| 182 | /** a ZeroR model in case no model can be built from the data */ |
---|
| 183 | protected Classifier m_ZeroR; |
---|
| 184 | |
---|
| 185 | /** |
---|
| 186 | * Returns a string describing classifier |
---|
| 187 | * @return a description suitable for |
---|
| 188 | * displaying in the explorer/experimenter gui |
---|
| 189 | */ |
---|
| 190 | public String globalInfo() { |
---|
| 191 | |
---|
| 192 | return "Class for performing additive logistic regression. \n" |
---|
| 193 | + "This class performs classification using a regression scheme as the " |
---|
| 194 | + "base learner, and can handle multi-class problems. For more " |
---|
| 195 | + "information, see\n\n" |
---|
| 196 | + getTechnicalInformation().toString() + "\n\n" |
---|
| 197 | + "Can do efficient internal cross-validation to determine " |
---|
| 198 | + "appropriate number of iterations."; |
---|
| 199 | } |
---|
| 200 | |
---|
| 201 | /** |
---|
| 202 | * Constructor. |
---|
| 203 | */ |
---|
| 204 | public LogitBoost() { |
---|
| 205 | |
---|
| 206 | m_Classifier = new weka.classifiers.trees.DecisionStump(); |
---|
| 207 | } |
---|
| 208 | |
---|
| 209 | /** |
---|
| 210 | * Returns an instance of a TechnicalInformation object, containing |
---|
| 211 | * detailed information about the technical background of this class, |
---|
| 212 | * e.g., paper reference or book this class is based on. |
---|
| 213 | * |
---|
| 214 | * @return the technical information about this class |
---|
| 215 | */ |
---|
| 216 | public TechnicalInformation getTechnicalInformation() { |
---|
| 217 | TechnicalInformation result; |
---|
| 218 | |
---|
| 219 | result = new TechnicalInformation(Type.TECHREPORT); |
---|
| 220 | result.setValue(Field.AUTHOR, "J. Friedman and T. Hastie and R. Tibshirani"); |
---|
| 221 | result.setValue(Field.YEAR, "1998"); |
---|
| 222 | result.setValue(Field.TITLE, "Additive Logistic Regression: a Statistical View of Boosting"); |
---|
| 223 | result.setValue(Field.ADDRESS, "Stanford University"); |
---|
| 224 | result.setValue(Field.PS, "http://www-stat.stanford.edu/~jhf/ftp/boost.ps"); |
---|
| 225 | |
---|
| 226 | return result; |
---|
| 227 | } |
---|
| 228 | |
---|
| 229 | /** |
---|
| 230 | * String describing default classifier. |
---|
| 231 | * |
---|
| 232 | * @return the default classifier classname |
---|
| 233 | */ |
---|
| 234 | protected String defaultClassifierString() { |
---|
| 235 | |
---|
| 236 | return "weka.classifiers.trees.DecisionStump"; |
---|
| 237 | } |
---|
| 238 | |
---|
| 239 | /** |
---|
| 240 | * Select only instances with weights that contribute to |
---|
| 241 | * the specified quantile of the weight distribution |
---|
| 242 | * |
---|
| 243 | * @param data the input instances |
---|
| 244 | * @param quantile the specified quantile eg 0.9 to select |
---|
| 245 | * 90% of the weight mass |
---|
| 246 | * @return the selected instances |
---|
| 247 | */ |
---|
| 248 | protected Instances selectWeightQuantile(Instances data, double quantile) { |
---|
| 249 | |
---|
| 250 | int numInstances = data.numInstances(); |
---|
| 251 | Instances trainData = new Instances(data, numInstances); |
---|
| 252 | double [] weights = new double [numInstances]; |
---|
| 253 | |
---|
| 254 | double sumOfWeights = 0; |
---|
| 255 | for (int i = 0; i < numInstances; i++) { |
---|
| 256 | weights[i] = data.instance(i).weight(); |
---|
| 257 | sumOfWeights += weights[i]; |
---|
| 258 | } |
---|
| 259 | double weightMassToSelect = sumOfWeights * quantile; |
---|
| 260 | int [] sortedIndices = Utils.sort(weights); |
---|
| 261 | |
---|
| 262 | // Select the instances |
---|
| 263 | sumOfWeights = 0; |
---|
| 264 | for (int i = numInstances-1; i >= 0; i--) { |
---|
| 265 | Instance instance = (Instance)data.instance(sortedIndices[i]).copy(); |
---|
| 266 | trainData.add(instance); |
---|
| 267 | sumOfWeights += weights[sortedIndices[i]]; |
---|
| 268 | if ((sumOfWeights > weightMassToSelect) && |
---|
| 269 | (i > 0) && |
---|
| 270 | (weights[sortedIndices[i]] != weights[sortedIndices[i-1]])) { |
---|
| 271 | break; |
---|
| 272 | } |
---|
| 273 | } |
---|
| 274 | if (m_Debug) { |
---|
| 275 | System.err.println("Selected " + trainData.numInstances() |
---|
| 276 | + " out of " + numInstances); |
---|
| 277 | } |
---|
| 278 | return trainData; |
---|
| 279 | } |
---|
| 280 | |
---|
| 281 | /** |
---|
| 282 | * Returns an enumeration describing the available options. |
---|
| 283 | * |
---|
| 284 | * @return an enumeration of all the available options. |
---|
| 285 | */ |
---|
| 286 | public Enumeration listOptions() { |
---|
| 287 | |
---|
| 288 | Vector newVector = new Vector(6); |
---|
| 289 | |
---|
| 290 | newVector.addElement(new Option( |
---|
| 291 | "\tUse resampling instead of reweighting for boosting.", |
---|
| 292 | "Q", 0, "-Q")); |
---|
| 293 | newVector.addElement(new Option( |
---|
| 294 | "\tPercentage of weight mass to base training on.\n" |
---|
| 295 | +"\t(default 100, reduce to around 90 speed up)", |
---|
| 296 | "P", 1, "-P <percent>")); |
---|
| 297 | newVector.addElement(new Option( |
---|
| 298 | "\tNumber of folds for internal cross-validation.\n" |
---|
| 299 | +"\t(default 0 -- no cross-validation)", |
---|
| 300 | "F", 1, "-F <num>")); |
---|
| 301 | newVector.addElement(new Option( |
---|
| 302 | "\tNumber of runs for internal cross-validation.\n" |
---|
| 303 | +"\t(default 1)", |
---|
| 304 | "R", 1, "-R <num>")); |
---|
| 305 | newVector.addElement(new Option( |
---|
| 306 | "\tThreshold on the improvement of the likelihood.\n" |
---|
| 307 | +"\t(default -Double.MAX_VALUE)", |
---|
| 308 | "L", 1, "-L <num>")); |
---|
| 309 | newVector.addElement(new Option( |
---|
| 310 | "\tShrinkage parameter.\n" |
---|
| 311 | +"\t(default 1)", |
---|
| 312 | "H", 1, "-H <num>")); |
---|
| 313 | |
---|
| 314 | Enumeration enu = super.listOptions(); |
---|
| 315 | while (enu.hasMoreElements()) { |
---|
| 316 | newVector.addElement(enu.nextElement()); |
---|
| 317 | } |
---|
| 318 | return newVector.elements(); |
---|
| 319 | } |
---|
| 320 | |
---|
| 321 | |
---|
| 322 | /** |
---|
| 323 | * Parses a given list of options. <p/> |
---|
| 324 | * |
---|
| 325 | <!-- options-start --> |
---|
| 326 | * Valid options are: <p/> |
---|
| 327 | * |
---|
| 328 | * <pre> -Q |
---|
| 329 | * Use resampling instead of reweighting for boosting.</pre> |
---|
| 330 | * |
---|
| 331 | * <pre> -P <percent> |
---|
| 332 | * Percentage of weight mass to base training on. |
---|
| 333 | * (default 100, reduce to around 90 speed up)</pre> |
---|
| 334 | * |
---|
| 335 | * <pre> -F <num> |
---|
| 336 | * Number of folds for internal cross-validation. |
---|
| 337 | * (default 0 -- no cross-validation)</pre> |
---|
| 338 | * |
---|
| 339 | * <pre> -R <num> |
---|
| 340 | * Number of runs for internal cross-validation. |
---|
| 341 | * (default 1)</pre> |
---|
| 342 | * |
---|
| 343 | * <pre> -L <num> |
---|
| 344 | * Threshold on the improvement of the likelihood. |
---|
| 345 | * (default -Double.MAX_VALUE)</pre> |
---|
| 346 | * |
---|
| 347 | * <pre> -H <num> |
---|
| 348 | * Shrinkage parameter. |
---|
| 349 | * (default 1)</pre> |
---|
| 350 | * |
---|
| 351 | * <pre> -S <num> |
---|
| 352 | * Random number seed. |
---|
| 353 | * (default 1)</pre> |
---|
| 354 | * |
---|
| 355 | * <pre> -I <num> |
---|
| 356 | * Number of iterations. |
---|
| 357 | * (default 10)</pre> |
---|
| 358 | * |
---|
| 359 | * <pre> -D |
---|
| 360 | * If set, classifier is run in debug mode and |
---|
| 361 | * may output additional info to the console</pre> |
---|
| 362 | * |
---|
| 363 | * <pre> -W |
---|
| 364 | * Full name of base classifier. |
---|
| 365 | * (default: weka.classifiers.trees.DecisionStump)</pre> |
---|
| 366 | * |
---|
| 367 | * <pre> |
---|
| 368 | * Options specific to classifier weka.classifiers.trees.DecisionStump: |
---|
| 369 | * </pre> |
---|
| 370 | * |
---|
| 371 | * <pre> -D |
---|
| 372 | * If set, classifier is run in debug mode and |
---|
| 373 | * may output additional info to the console</pre> |
---|
| 374 | * |
---|
| 375 | <!-- options-end --> |
---|
| 376 | * |
---|
| 377 | * Options after -- are passed to the designated learner.<p> |
---|
| 378 | * |
---|
| 379 | * @param options the list of options as an array of strings |
---|
| 380 | * @throws Exception if an option is not supported |
---|
| 381 | */ |
---|
| 382 | public void setOptions(String[] options) throws Exception { |
---|
| 383 | |
---|
| 384 | String numFolds = Utils.getOption('F', options); |
---|
| 385 | if (numFolds.length() != 0) { |
---|
| 386 | setNumFolds(Integer.parseInt(numFolds)); |
---|
| 387 | } else { |
---|
| 388 | setNumFolds(0); |
---|
| 389 | } |
---|
| 390 | |
---|
| 391 | String numRuns = Utils.getOption('R', options); |
---|
| 392 | if (numRuns.length() != 0) { |
---|
| 393 | setNumRuns(Integer.parseInt(numRuns)); |
---|
| 394 | } else { |
---|
| 395 | setNumRuns(1); |
---|
| 396 | } |
---|
| 397 | |
---|
| 398 | String thresholdString = Utils.getOption('P', options); |
---|
| 399 | if (thresholdString.length() != 0) { |
---|
| 400 | setWeightThreshold(Integer.parseInt(thresholdString)); |
---|
| 401 | } else { |
---|
| 402 | setWeightThreshold(100); |
---|
| 403 | } |
---|
| 404 | |
---|
| 405 | String precisionString = Utils.getOption('L', options); |
---|
| 406 | if (precisionString.length() != 0) { |
---|
| 407 | setLikelihoodThreshold(new Double(precisionString). |
---|
| 408 | doubleValue()); |
---|
| 409 | } else { |
---|
| 410 | setLikelihoodThreshold(-Double.MAX_VALUE); |
---|
| 411 | } |
---|
| 412 | |
---|
| 413 | String shrinkageString = Utils.getOption('H', options); |
---|
| 414 | if (shrinkageString.length() != 0) { |
---|
| 415 | setShrinkage(new Double(shrinkageString). |
---|
| 416 | doubleValue()); |
---|
| 417 | } else { |
---|
| 418 | setShrinkage(1.0); |
---|
| 419 | } |
---|
| 420 | |
---|
| 421 | setUseResampling(Utils.getFlag('Q', options)); |
---|
| 422 | if (m_UseResampling && (thresholdString.length() != 0)) { |
---|
| 423 | throw new Exception("Weight pruning with resampling"+ |
---|
| 424 | "not allowed."); |
---|
| 425 | } |
---|
| 426 | |
---|
| 427 | super.setOptions(options); |
---|
| 428 | } |
---|
| 429 | |
---|
| 430 | /** |
---|
| 431 | * Gets the current settings of the Classifier. |
---|
| 432 | * |
---|
| 433 | * @return an array of strings suitable for passing to setOptions |
---|
| 434 | */ |
---|
| 435 | public String [] getOptions() { |
---|
| 436 | |
---|
| 437 | String [] superOptions = super.getOptions(); |
---|
| 438 | String [] options = new String [superOptions.length + 10]; |
---|
| 439 | |
---|
| 440 | int current = 0; |
---|
| 441 | if (getUseResampling()) { |
---|
| 442 | options[current++] = "-Q"; |
---|
| 443 | } else { |
---|
| 444 | options[current++] = "-P"; |
---|
| 445 | options[current++] = "" + getWeightThreshold(); |
---|
| 446 | } |
---|
| 447 | options[current++] = "-F"; options[current++] = "" + getNumFolds(); |
---|
| 448 | options[current++] = "-R"; options[current++] = "" + getNumRuns(); |
---|
| 449 | options[current++] = "-L"; options[current++] = "" + getLikelihoodThreshold(); |
---|
| 450 | options[current++] = "-H"; options[current++] = "" + getShrinkage(); |
---|
| 451 | |
---|
| 452 | System.arraycopy(superOptions, 0, options, current, |
---|
| 453 | superOptions.length); |
---|
| 454 | current += superOptions.length; |
---|
| 455 | while (current < options.length) { |
---|
| 456 | options[current++] = ""; |
---|
| 457 | } |
---|
| 458 | return options; |
---|
| 459 | } |
---|
| 460 | |
---|
| 461 | /** |
---|
| 462 | * Returns the tip text for this property |
---|
| 463 | * @return tip text for this property suitable for |
---|
| 464 | * displaying in the explorer/experimenter gui |
---|
| 465 | */ |
---|
| 466 | public String shrinkageTipText() { |
---|
| 467 | return "Shrinkage parameter (use small value like 0.1 to reduce " |
---|
| 468 | + "overfitting)."; |
---|
| 469 | } |
---|
| 470 | |
---|
| 471 | /** |
---|
| 472 | * Get the value of Shrinkage. |
---|
| 473 | * |
---|
| 474 | * @return Value of Shrinkage. |
---|
| 475 | */ |
---|
| 476 | public double getShrinkage() { |
---|
| 477 | |
---|
| 478 | return m_Shrinkage; |
---|
| 479 | } |
---|
| 480 | |
---|
| 481 | /** |
---|
| 482 | * Set the value of Shrinkage. |
---|
| 483 | * |
---|
| 484 | * @param newShrinkage Value to assign to Shrinkage. |
---|
| 485 | */ |
---|
| 486 | public void setShrinkage(double newShrinkage) { |
---|
| 487 | |
---|
| 488 | m_Shrinkage = newShrinkage; |
---|
| 489 | } |
---|
| 490 | |
---|
| 491 | /** |
---|
| 492 | * Returns the tip text for this property |
---|
| 493 | * @return tip text for this property suitable for |
---|
| 494 | * displaying in the explorer/experimenter gui |
---|
| 495 | */ |
---|
| 496 | public String likelihoodThresholdTipText() { |
---|
| 497 | return "Threshold on improvement in likelihood."; |
---|
| 498 | } |
---|
| 499 | |
---|
| 500 | /** |
---|
| 501 | * Get the value of Precision. |
---|
| 502 | * |
---|
| 503 | * @return Value of Precision. |
---|
| 504 | */ |
---|
| 505 | public double getLikelihoodThreshold() { |
---|
| 506 | |
---|
| 507 | return m_Precision; |
---|
| 508 | } |
---|
| 509 | |
---|
| 510 | /** |
---|
| 511 | * Set the value of Precision. |
---|
| 512 | * |
---|
| 513 | * @param newPrecision Value to assign to Precision. |
---|
| 514 | */ |
---|
| 515 | public void setLikelihoodThreshold(double newPrecision) { |
---|
| 516 | |
---|
| 517 | m_Precision = newPrecision; |
---|
| 518 | } |
---|
| 519 | |
---|
| 520 | /** |
---|
| 521 | * Returns the tip text for this property |
---|
| 522 | * @return tip text for this property suitable for |
---|
| 523 | * displaying in the explorer/experimenter gui |
---|
| 524 | */ |
---|
| 525 | public String numRunsTipText() { |
---|
| 526 | return "Number of runs for internal cross-validation."; |
---|
| 527 | } |
---|
| 528 | |
---|
| 529 | /** |
---|
| 530 | * Get the value of NumRuns. |
---|
| 531 | * |
---|
| 532 | * @return Value of NumRuns. |
---|
| 533 | */ |
---|
| 534 | public int getNumRuns() { |
---|
| 535 | |
---|
| 536 | return m_NumRuns; |
---|
| 537 | } |
---|
| 538 | |
---|
| 539 | /** |
---|
| 540 | * Set the value of NumRuns. |
---|
| 541 | * |
---|
| 542 | * @param newNumRuns Value to assign to NumRuns. |
---|
| 543 | */ |
---|
| 544 | public void setNumRuns(int newNumRuns) { |
---|
| 545 | |
---|
| 546 | m_NumRuns = newNumRuns; |
---|
| 547 | } |
---|
| 548 | |
---|
| 549 | /** |
---|
| 550 | * Returns the tip text for this property |
---|
| 551 | * @return tip text for this property suitable for |
---|
| 552 | * displaying in the explorer/experimenter gui |
---|
| 553 | */ |
---|
| 554 | public String numFoldsTipText() { |
---|
| 555 | return "Number of folds for internal cross-validation (default 0 " |
---|
| 556 | + "means no cross-validation is performed)."; |
---|
| 557 | } |
---|
| 558 | |
---|
| 559 | /** |
---|
| 560 | * Get the value of NumFolds. |
---|
| 561 | * |
---|
| 562 | * @return Value of NumFolds. |
---|
| 563 | */ |
---|
| 564 | public int getNumFolds() { |
---|
| 565 | |
---|
| 566 | return m_NumFolds; |
---|
| 567 | } |
---|
| 568 | |
---|
| 569 | /** |
---|
| 570 | * Set the value of NumFolds. |
---|
| 571 | * |
---|
| 572 | * @param newNumFolds Value to assign to NumFolds. |
---|
| 573 | */ |
---|
| 574 | public void setNumFolds(int newNumFolds) { |
---|
| 575 | |
---|
| 576 | m_NumFolds = newNumFolds; |
---|
| 577 | } |
---|
| 578 | |
---|
| 579 | /** |
---|
| 580 | * Returns the tip text for this property |
---|
| 581 | * @return tip text for this property suitable for |
---|
| 582 | * displaying in the explorer/experimenter gui |
---|
| 583 | */ |
---|
| 584 | public String useResamplingTipText() { |
---|
| 585 | return "Whether resampling is used instead of reweighting."; |
---|
| 586 | } |
---|
| 587 | |
---|
| 588 | /** |
---|
| 589 | * Set resampling mode |
---|
| 590 | * |
---|
| 591 | * @param r true if resampling should be done |
---|
| 592 | */ |
---|
| 593 | public void setUseResampling(boolean r) { |
---|
| 594 | |
---|
| 595 | m_UseResampling = r; |
---|
| 596 | } |
---|
| 597 | |
---|
| 598 | /** |
---|
| 599 | * Get whether resampling is turned on |
---|
| 600 | * |
---|
| 601 | * @return true if resampling output is on |
---|
| 602 | */ |
---|
| 603 | public boolean getUseResampling() { |
---|
| 604 | |
---|
| 605 | return m_UseResampling; |
---|
| 606 | } |
---|
| 607 | |
---|
| 608 | /** |
---|
| 609 | * Returns the tip text for this property |
---|
| 610 | * @return tip text for this property suitable for |
---|
| 611 | * displaying in the explorer/experimenter gui |
---|
| 612 | */ |
---|
| 613 | public String weightThresholdTipText() { |
---|
| 614 | return "Weight threshold for weight pruning (reduce to 90 " |
---|
| 615 | + "for speeding up learning process)."; |
---|
| 616 | } |
---|
| 617 | |
---|
| 618 | /** |
---|
| 619 | * Set weight thresholding |
---|
| 620 | * |
---|
| 621 | * @param threshold the percentage of weight mass used for training |
---|
| 622 | */ |
---|
| 623 | public void setWeightThreshold(int threshold) { |
---|
| 624 | |
---|
| 625 | m_WeightThreshold = threshold; |
---|
| 626 | } |
---|
| 627 | |
---|
| 628 | /** |
---|
| 629 | * Get the degree of weight thresholding |
---|
| 630 | * |
---|
| 631 | * @return the percentage of weight mass used for training |
---|
| 632 | */ |
---|
| 633 | public int getWeightThreshold() { |
---|
| 634 | |
---|
| 635 | return m_WeightThreshold; |
---|
| 636 | } |
---|
| 637 | |
---|
| 638 | /** |
---|
| 639 | * Returns default capabilities of the classifier. |
---|
| 640 | * |
---|
| 641 | * @return the capabilities of this classifier |
---|
| 642 | */ |
---|
| 643 | public Capabilities getCapabilities() { |
---|
| 644 | Capabilities result = super.getCapabilities(); |
---|
| 645 | |
---|
| 646 | // class |
---|
| 647 | result.disableAllClasses(); |
---|
| 648 | result.disableAllClassDependencies(); |
---|
| 649 | result.enable(Capability.NOMINAL_CLASS); |
---|
| 650 | |
---|
| 651 | return result; |
---|
| 652 | } |
---|
| 653 | |
---|
| 654 | /** |
---|
| 655 | * Builds the boosted classifier |
---|
| 656 | * |
---|
| 657 | * @param data the data to train the classifier with |
---|
| 658 | * @throws Exception if building fails, e.g., can't handle data |
---|
| 659 | */ |
---|
| 660 | public void buildClassifier(Instances data) throws Exception { |
---|
| 661 | |
---|
| 662 | m_RandomInstance = new Random(m_Seed); |
---|
| 663 | int classIndex = data.classIndex(); |
---|
| 664 | |
---|
| 665 | if (m_Classifier == null) { |
---|
| 666 | throw new Exception("A base classifier has not been specified!"); |
---|
| 667 | } |
---|
| 668 | |
---|
| 669 | if (!(m_Classifier instanceof WeightedInstancesHandler) && |
---|
| 670 | !m_UseResampling) { |
---|
| 671 | m_UseResampling = true; |
---|
| 672 | } |
---|
| 673 | |
---|
| 674 | // can classifier handle the data? |
---|
| 675 | getCapabilities().testWithFail(data); |
---|
| 676 | |
---|
| 677 | if (m_Debug) { |
---|
| 678 | System.err.println("Creating copy of the training data"); |
---|
| 679 | } |
---|
| 680 | |
---|
| 681 | // remove instances with missing class |
---|
| 682 | data = new Instances(data); |
---|
| 683 | data.deleteWithMissingClass(); |
---|
| 684 | |
---|
| 685 | // only class? -> build ZeroR model |
---|
| 686 | if (data.numAttributes() == 1) { |
---|
| 687 | System.err.println( |
---|
| 688 | "Cannot build model (only class attribute present in data!), " |
---|
| 689 | + "using ZeroR model instead!"); |
---|
| 690 | m_ZeroR = new weka.classifiers.rules.ZeroR(); |
---|
| 691 | m_ZeroR.buildClassifier(data); |
---|
| 692 | return; |
---|
| 693 | } |
---|
| 694 | else { |
---|
| 695 | m_ZeroR = null; |
---|
| 696 | } |
---|
| 697 | |
---|
| 698 | m_NumClasses = data.numClasses(); |
---|
| 699 | m_ClassAttribute = data.classAttribute(); |
---|
| 700 | |
---|
| 701 | // Create the base classifiers |
---|
| 702 | if (m_Debug) { |
---|
| 703 | System.err.println("Creating base classifiers"); |
---|
| 704 | } |
---|
| 705 | m_Classifiers = new Classifier [m_NumClasses][]; |
---|
| 706 | for (int j = 0; j < m_NumClasses; j++) { |
---|
| 707 | m_Classifiers[j] = AbstractClassifier.makeCopies(m_Classifier, |
---|
| 708 | getNumIterations()); |
---|
| 709 | } |
---|
| 710 | |
---|
| 711 | // Do we want to select the appropriate number of iterations |
---|
| 712 | // using cross-validation? |
---|
| 713 | int bestNumIterations = getNumIterations(); |
---|
| 714 | if (m_NumFolds > 1) { |
---|
| 715 | if (m_Debug) { |
---|
| 716 | System.err.println("Processing first fold."); |
---|
| 717 | } |
---|
| 718 | |
---|
| 719 | // Array for storing the results |
---|
| 720 | double[] results = new double[getNumIterations()]; |
---|
| 721 | |
---|
| 722 | // Iterate throught the cv-runs |
---|
| 723 | for (int r = 0; r < m_NumRuns; r++) { |
---|
| 724 | |
---|
| 725 | // Stratify the data |
---|
| 726 | data.randomize(m_RandomInstance); |
---|
| 727 | data.stratify(m_NumFolds); |
---|
| 728 | |
---|
| 729 | // Perform the cross-validation |
---|
| 730 | for (int i = 0; i < m_NumFolds; i++) { |
---|
| 731 | |
---|
| 732 | // Get train and test folds |
---|
| 733 | Instances train = data.trainCV(m_NumFolds, i, m_RandomInstance); |
---|
| 734 | Instances test = data.testCV(m_NumFolds, i); |
---|
| 735 | |
---|
| 736 | // Make class numeric |
---|
| 737 | Instances trainN = new Instances(train); |
---|
| 738 | trainN.setClassIndex(-1); |
---|
| 739 | trainN.deleteAttributeAt(classIndex); |
---|
| 740 | trainN.insertAttributeAt(new Attribute("'pseudo class'"), classIndex); |
---|
| 741 | trainN.setClassIndex(classIndex); |
---|
| 742 | m_NumericClassData = new Instances(trainN, 0); |
---|
| 743 | |
---|
| 744 | // Get class values |
---|
| 745 | int numInstances = train.numInstances(); |
---|
| 746 | double [][] trainFs = new double [numInstances][m_NumClasses]; |
---|
| 747 | double [][] trainYs = new double [numInstances][m_NumClasses]; |
---|
| 748 | for (int j = 0; j < m_NumClasses; j++) { |
---|
| 749 | for (int k = 0; k < numInstances; k++) { |
---|
| 750 | trainYs[k][j] = (train.instance(k).classValue() == j) ? |
---|
| 751 | 1.0 - m_Offset: 0.0 + (m_Offset / (double)m_NumClasses); |
---|
| 752 | } |
---|
| 753 | } |
---|
| 754 | |
---|
| 755 | // Perform iterations |
---|
| 756 | double[][] probs = initialProbs(numInstances); |
---|
| 757 | m_NumGenerated = 0; |
---|
| 758 | double sumOfWeights = train.sumOfWeights(); |
---|
| 759 | for (int j = 0; j < getNumIterations(); j++) { |
---|
| 760 | performIteration(trainYs, trainFs, probs, trainN, sumOfWeights); |
---|
| 761 | Evaluation eval = new Evaluation(train); |
---|
| 762 | eval.evaluateModel(this, test); |
---|
| 763 | results[j] += eval.correct(); |
---|
| 764 | } |
---|
| 765 | } |
---|
| 766 | } |
---|
| 767 | |
---|
| 768 | // Find the number of iterations with the lowest error |
---|
| 769 | double bestResult = -Double.MAX_VALUE; |
---|
| 770 | for (int j = 0; j < getNumIterations(); j++) { |
---|
| 771 | if (results[j] > bestResult) { |
---|
| 772 | bestResult = results[j]; |
---|
| 773 | bestNumIterations = j; |
---|
| 774 | } |
---|
| 775 | } |
---|
| 776 | if (m_Debug) { |
---|
| 777 | System.err.println("Best result for " + |
---|
| 778 | bestNumIterations + " iterations: " + |
---|
| 779 | bestResult); |
---|
| 780 | } |
---|
| 781 | } |
---|
| 782 | |
---|
| 783 | // Build classifier on all the data |
---|
| 784 | int numInstances = data.numInstances(); |
---|
| 785 | double [][] trainFs = new double [numInstances][m_NumClasses]; |
---|
| 786 | double [][] trainYs = new double [numInstances][m_NumClasses]; |
---|
| 787 | for (int j = 0; j < m_NumClasses; j++) { |
---|
| 788 | for (int i = 0, k = 0; i < numInstances; i++, k++) { |
---|
| 789 | trainYs[i][j] = (data.instance(k).classValue() == j) ? |
---|
| 790 | 1.0 - m_Offset: 0.0 + (m_Offset / (double)m_NumClasses); |
---|
| 791 | } |
---|
| 792 | } |
---|
| 793 | |
---|
| 794 | // Make class numeric |
---|
| 795 | data.setClassIndex(-1); |
---|
| 796 | data.deleteAttributeAt(classIndex); |
---|
| 797 | data.insertAttributeAt(new Attribute("'pseudo class'"), classIndex); |
---|
| 798 | data.setClassIndex(classIndex); |
---|
| 799 | m_NumericClassData = new Instances(data, 0); |
---|
| 800 | |
---|
| 801 | // Perform iterations |
---|
| 802 | double[][] probs = initialProbs(numInstances); |
---|
| 803 | double logLikelihood = logLikelihood(trainYs, probs); |
---|
| 804 | m_NumGenerated = 0; |
---|
| 805 | if (m_Debug) { |
---|
| 806 | System.err.println("Avg. log-likelihood: " + logLikelihood); |
---|
| 807 | } |
---|
| 808 | double sumOfWeights = data.sumOfWeights(); |
---|
| 809 | for (int j = 0; j < bestNumIterations; j++) { |
---|
| 810 | double previousLoglikelihood = logLikelihood; |
---|
| 811 | performIteration(trainYs, trainFs, probs, data, sumOfWeights); |
---|
| 812 | logLikelihood = logLikelihood(trainYs, probs); |
---|
| 813 | if (m_Debug) { |
---|
| 814 | System.err.println("Avg. log-likelihood: " + logLikelihood); |
---|
| 815 | } |
---|
| 816 | if (Math.abs(previousLoglikelihood - logLikelihood) < m_Precision) { |
---|
| 817 | return; |
---|
| 818 | } |
---|
| 819 | } |
---|
| 820 | } |
---|
| 821 | |
---|
| 822 | /** |
---|
| 823 | * Gets the intial class probabilities. |
---|
| 824 | * |
---|
| 825 | * @param numInstances the number of instances |
---|
| 826 | * @return the initial class probabilities |
---|
| 827 | */ |
---|
| 828 | private double[][] initialProbs(int numInstances) { |
---|
| 829 | |
---|
| 830 | double[][] probs = new double[numInstances][m_NumClasses]; |
---|
| 831 | for (int i = 0; i < numInstances; i++) { |
---|
| 832 | for (int j = 0 ; j < m_NumClasses; j++) { |
---|
| 833 | probs[i][j] = 1.0 / m_NumClasses; |
---|
| 834 | } |
---|
| 835 | } |
---|
| 836 | return probs; |
---|
| 837 | } |
---|
| 838 | |
---|
| 839 | /** |
---|
| 840 | * Computes loglikelihood given class values |
---|
| 841 | * and estimated probablities. |
---|
| 842 | * |
---|
| 843 | * @param trainYs class values |
---|
| 844 | * @param probs estimated probabilities |
---|
| 845 | * @return the computed loglikelihood |
---|
| 846 | */ |
---|
| 847 | private double logLikelihood(double[][] trainYs, double[][] probs) { |
---|
| 848 | |
---|
| 849 | double logLikelihood = 0; |
---|
| 850 | for (int i = 0; i < trainYs.length; i++) { |
---|
| 851 | for (int j = 0; j < m_NumClasses; j++) { |
---|
| 852 | if (trainYs[i][j] == 1.0 - m_Offset) { |
---|
| 853 | logLikelihood -= Math.log(probs[i][j]); |
---|
| 854 | } |
---|
| 855 | } |
---|
| 856 | } |
---|
| 857 | return logLikelihood / (double)trainYs.length; |
---|
| 858 | } |
---|
| 859 | |
---|
| 860 | /** |
---|
| 861 | * Performs one boosting iteration. |
---|
| 862 | * |
---|
| 863 | * @param trainYs class values |
---|
| 864 | * @param trainFs F scores |
---|
| 865 | * @param probs probabilities |
---|
| 866 | * @param data the data to run the iteration on |
---|
| 867 | * @param origSumOfWeights the original sum of weights |
---|
| 868 | * @throws Exception in case base classifiers run into problems |
---|
| 869 | */ |
---|
| 870 | private void performIteration(double[][] trainYs, |
---|
| 871 | double[][] trainFs, |
---|
| 872 | double[][] probs, |
---|
| 873 | Instances data, |
---|
| 874 | double origSumOfWeights) throws Exception { |
---|
| 875 | |
---|
| 876 | if (m_Debug) { |
---|
| 877 | System.err.println("Training classifier " + (m_NumGenerated + 1)); |
---|
| 878 | } |
---|
| 879 | |
---|
| 880 | // Build the new models |
---|
| 881 | for (int j = 0; j < m_NumClasses; j++) { |
---|
| 882 | if (m_Debug) { |
---|
| 883 | System.err.println("\t...for class " + (j + 1) |
---|
| 884 | + " (" + m_ClassAttribute.name() |
---|
| 885 | + "=" + m_ClassAttribute.value(j) + ")"); |
---|
| 886 | } |
---|
| 887 | |
---|
| 888 | // Make copy because we want to save the weights |
---|
| 889 | Instances boostData = new Instances(data); |
---|
| 890 | |
---|
| 891 | // Set instance pseudoclass and weights |
---|
| 892 | for (int i = 0; i < probs.length; i++) { |
---|
| 893 | |
---|
| 894 | // Compute response and weight |
---|
| 895 | double p = probs[i][j]; |
---|
| 896 | double z, actual = trainYs[i][j]; |
---|
| 897 | if (actual == 1 - m_Offset) { |
---|
| 898 | z = 1.0 / p; |
---|
| 899 | if (z > Z_MAX) { // threshold |
---|
| 900 | z = Z_MAX; |
---|
| 901 | } |
---|
| 902 | } else { |
---|
| 903 | z = -1.0 / (1.0 - p); |
---|
| 904 | if (z < -Z_MAX) { // threshold |
---|
| 905 | z = -Z_MAX; |
---|
| 906 | } |
---|
| 907 | } |
---|
| 908 | double w = (actual - p) / z; |
---|
| 909 | |
---|
| 910 | // Set values for instance |
---|
| 911 | Instance current = boostData.instance(i); |
---|
| 912 | current.setValue(boostData.classIndex(), z); |
---|
| 913 | current.setWeight(current.weight() * w); |
---|
| 914 | } |
---|
| 915 | |
---|
| 916 | // Scale the weights (helps with some base learners) |
---|
| 917 | double sumOfWeights = boostData.sumOfWeights(); |
---|
| 918 | double scalingFactor = (double)origSumOfWeights / sumOfWeights; |
---|
| 919 | for (int i = 0; i < probs.length; i++) { |
---|
| 920 | Instance current = boostData.instance(i); |
---|
| 921 | current.setWeight(current.weight() * scalingFactor); |
---|
| 922 | } |
---|
| 923 | |
---|
| 924 | // Select instances to train the classifier on |
---|
| 925 | Instances trainData = boostData; |
---|
| 926 | if (m_WeightThreshold < 100) { |
---|
| 927 | trainData = selectWeightQuantile(boostData, |
---|
| 928 | (double)m_WeightThreshold / 100); |
---|
| 929 | } else { |
---|
| 930 | if (m_UseResampling) { |
---|
| 931 | double[] weights = new double[boostData.numInstances()]; |
---|
| 932 | for (int kk = 0; kk < weights.length; kk++) { |
---|
| 933 | weights[kk] = boostData.instance(kk).weight(); |
---|
| 934 | } |
---|
| 935 | trainData = boostData.resampleWithWeights(m_RandomInstance, |
---|
| 936 | weights); |
---|
| 937 | } |
---|
| 938 | } |
---|
| 939 | |
---|
| 940 | // Build the classifier |
---|
| 941 | m_Classifiers[j][m_NumGenerated].buildClassifier(trainData); |
---|
| 942 | } |
---|
| 943 | |
---|
| 944 | // Evaluate / increment trainFs from the classifier |
---|
| 945 | for (int i = 0; i < trainFs.length; i++) { |
---|
| 946 | double [] pred = new double [m_NumClasses]; |
---|
| 947 | double predSum = 0; |
---|
| 948 | for (int j = 0; j < m_NumClasses; j++) { |
---|
| 949 | pred[j] = m_Shrinkage * m_Classifiers[j][m_NumGenerated] |
---|
| 950 | .classifyInstance(data.instance(i)); |
---|
| 951 | predSum += pred[j]; |
---|
| 952 | } |
---|
| 953 | predSum /= m_NumClasses; |
---|
| 954 | for (int j = 0; j < m_NumClasses; j++) { |
---|
| 955 | trainFs[i][j] += (pred[j] - predSum) * (m_NumClasses - 1) |
---|
| 956 | / m_NumClasses; |
---|
| 957 | } |
---|
| 958 | } |
---|
| 959 | m_NumGenerated++; |
---|
| 960 | |
---|
| 961 | // Compute the current probability estimates |
---|
| 962 | for (int i = 0; i < trainYs.length; i++) { |
---|
| 963 | probs[i] = probs(trainFs[i]); |
---|
| 964 | } |
---|
| 965 | } |
---|
| 966 | |
---|
| 967 | /** |
---|
| 968 | * Returns the array of classifiers that have been built. |
---|
| 969 | * |
---|
| 970 | * @return the built classifiers |
---|
| 971 | */ |
---|
| 972 | public Classifier[][] classifiers() { |
---|
| 973 | |
---|
| 974 | Classifier[][] classifiers = |
---|
| 975 | new Classifier[m_NumClasses][m_NumGenerated]; |
---|
| 976 | for (int j = 0; j < m_NumClasses; j++) { |
---|
| 977 | for (int i = 0; i < m_NumGenerated; i++) { |
---|
| 978 | classifiers[j][i] = m_Classifiers[j][i]; |
---|
| 979 | } |
---|
| 980 | } |
---|
| 981 | return classifiers; |
---|
| 982 | } |
---|
| 983 | |
---|
| 984 | /** |
---|
| 985 | * Computes probabilities from F scores |
---|
| 986 | * |
---|
| 987 | * @param Fs the F scores |
---|
| 988 | * @return the computed probabilities |
---|
| 989 | */ |
---|
| 990 | private double[] probs(double[] Fs) { |
---|
| 991 | |
---|
| 992 | double maxF = -Double.MAX_VALUE; |
---|
| 993 | for (int i = 0; i < Fs.length; i++) { |
---|
| 994 | if (Fs[i] > maxF) { |
---|
| 995 | maxF = Fs[i]; |
---|
| 996 | } |
---|
| 997 | } |
---|
| 998 | double sum = 0; |
---|
| 999 | double[] probs = new double[Fs.length]; |
---|
| 1000 | for (int i = 0; i < Fs.length; i++) { |
---|
| 1001 | probs[i] = Math.exp(Fs[i] - maxF); |
---|
| 1002 | sum += probs[i]; |
---|
| 1003 | } |
---|
| 1004 | Utils.normalize(probs, sum); |
---|
| 1005 | return probs; |
---|
| 1006 | } |
---|
| 1007 | |
---|
| 1008 | /** |
---|
| 1009 | * Calculates the class membership probabilities for the given test instance. |
---|
| 1010 | * |
---|
| 1011 | * @param instance the instance to be classified |
---|
| 1012 | * @return predicted class probability distribution |
---|
| 1013 | * @throws Exception if instance could not be classified |
---|
| 1014 | * successfully |
---|
| 1015 | */ |
---|
| 1016 | public double [] distributionForInstance(Instance instance) |
---|
| 1017 | throws Exception { |
---|
| 1018 | |
---|
| 1019 | // default model? |
---|
| 1020 | if (m_ZeroR != null) { |
---|
| 1021 | return m_ZeroR.distributionForInstance(instance); |
---|
| 1022 | } |
---|
| 1023 | |
---|
| 1024 | instance = (Instance)instance.copy(); |
---|
| 1025 | instance.setDataset(m_NumericClassData); |
---|
| 1026 | double [] pred = new double [m_NumClasses]; |
---|
| 1027 | double [] Fs = new double [m_NumClasses]; |
---|
| 1028 | for (int i = 0; i < m_NumGenerated; i++) { |
---|
| 1029 | double predSum = 0; |
---|
| 1030 | for (int j = 0; j < m_NumClasses; j++) { |
---|
| 1031 | pred[j] = m_Shrinkage * m_Classifiers[j][i].classifyInstance(instance); |
---|
| 1032 | predSum += pred[j]; |
---|
| 1033 | } |
---|
| 1034 | predSum /= m_NumClasses; |
---|
| 1035 | for (int j = 0; j < m_NumClasses; j++) { |
---|
| 1036 | Fs[j] += (pred[j] - predSum) * (m_NumClasses - 1) |
---|
| 1037 | / m_NumClasses; |
---|
| 1038 | } |
---|
| 1039 | } |
---|
| 1040 | |
---|
| 1041 | return probs(Fs); |
---|
| 1042 | } |
---|
| 1043 | |
---|
| 1044 | /** |
---|
| 1045 | * Returns the boosted model as Java source code. |
---|
| 1046 | * |
---|
| 1047 | * @param className the classname in the generated code |
---|
| 1048 | * @return the tree as Java source code |
---|
| 1049 | * @throws Exception if something goes wrong |
---|
| 1050 | */ |
---|
| 1051 | public String toSource(String className) throws Exception { |
---|
| 1052 | |
---|
| 1053 | if (m_NumGenerated == 0) { |
---|
| 1054 | throw new Exception("No model built yet"); |
---|
| 1055 | } |
---|
| 1056 | if (!(m_Classifiers[0][0] instanceof Sourcable)) { |
---|
| 1057 | throw new Exception("Base learner " + m_Classifier.getClass().getName() |
---|
| 1058 | + " is not Sourcable"); |
---|
| 1059 | } |
---|
| 1060 | |
---|
| 1061 | StringBuffer text = new StringBuffer("class "); |
---|
| 1062 | text.append(className).append(" {\n\n"); |
---|
| 1063 | text.append(" private static double RtoP(double []R, int j) {\n"+ |
---|
| 1064 | " double Rcenter = 0;\n"+ |
---|
| 1065 | " for (int i = 0; i < R.length; i++) {\n"+ |
---|
| 1066 | " Rcenter += R[i];\n"+ |
---|
| 1067 | " }\n"+ |
---|
| 1068 | " Rcenter /= R.length;\n"+ |
---|
| 1069 | " double Rsum = 0;\n"+ |
---|
| 1070 | " for (int i = 0; i < R.length; i++) {\n"+ |
---|
| 1071 | " Rsum += Math.exp(R[i] - Rcenter);\n"+ |
---|
| 1072 | " }\n"+ |
---|
| 1073 | " return Math.exp(R[j]) / Rsum;\n"+ |
---|
| 1074 | " }\n\n"); |
---|
| 1075 | |
---|
| 1076 | text.append(" public static double classify(Object[] i) {\n" + |
---|
| 1077 | " double [] d = distribution(i);\n" + |
---|
| 1078 | " double maxV = d[0];\n" + |
---|
| 1079 | " int maxI = 0;\n"+ |
---|
| 1080 | " for (int j = 1; j < " + m_NumClasses + "; j++) {\n"+ |
---|
| 1081 | " if (d[j] > maxV) { maxV = d[j]; maxI = j; }\n"+ |
---|
| 1082 | " }\n return (double) maxI;\n }\n\n"); |
---|
| 1083 | |
---|
| 1084 | text.append(" public static double [] distribution(Object [] i) {\n"); |
---|
| 1085 | text.append(" double [] Fs = new double [" + m_NumClasses + "];\n"); |
---|
| 1086 | text.append(" double [] Fi = new double [" + m_NumClasses + "];\n"); |
---|
| 1087 | text.append(" double Fsum;\n"); |
---|
| 1088 | for (int i = 0; i < m_NumGenerated; i++) { |
---|
| 1089 | text.append(" Fsum = 0;\n"); |
---|
| 1090 | for (int j = 0; j < m_NumClasses; j++) { |
---|
| 1091 | text.append(" Fi[" + j + "] = " + className + '_' +j + '_' + i |
---|
| 1092 | + ".classify(i); Fsum += Fi[" + j + "];\n"); |
---|
| 1093 | } |
---|
| 1094 | text.append(" Fsum /= " + m_NumClasses + ";\n"); |
---|
| 1095 | text.append(" for (int j = 0; j < " + m_NumClasses + "; j++) {"); |
---|
| 1096 | text.append(" Fs[j] += (Fi[j] - Fsum) * " |
---|
| 1097 | + (m_NumClasses - 1) + " / " + m_NumClasses + "; }\n"); |
---|
| 1098 | } |
---|
| 1099 | |
---|
| 1100 | text.append(" double [] dist = new double [" + m_NumClasses + "];\n" + |
---|
| 1101 | " for (int j = 0; j < " + m_NumClasses + "; j++) {\n"+ |
---|
| 1102 | " dist[j] = RtoP(Fs, j);\n"+ |
---|
| 1103 | " }\n return dist;\n"); |
---|
| 1104 | text.append(" }\n}\n"); |
---|
| 1105 | |
---|
| 1106 | for (int i = 0; i < m_Classifiers.length; i++) { |
---|
| 1107 | for (int j = 0; j < m_Classifiers[i].length; j++) { |
---|
| 1108 | text.append(((Sourcable)m_Classifiers[i][j]) |
---|
| 1109 | .toSource(className + '_' + i + '_' + j)); |
---|
| 1110 | } |
---|
| 1111 | } |
---|
| 1112 | return text.toString(); |
---|
| 1113 | } |
---|
| 1114 | |
---|
| 1115 | /** |
---|
| 1116 | * Returns description of the boosted classifier. |
---|
| 1117 | * |
---|
| 1118 | * @return description of the boosted classifier as a string |
---|
| 1119 | */ |
---|
| 1120 | public String toString() { |
---|
| 1121 | |
---|
| 1122 | // only ZeroR model? |
---|
| 1123 | if (m_ZeroR != null) { |
---|
| 1124 | StringBuffer buf = new StringBuffer(); |
---|
| 1125 | buf.append(this.getClass().getName().replaceAll(".*\\.", "") + "\n"); |
---|
| 1126 | buf.append(this.getClass().getName().replaceAll(".*\\.", "").replaceAll(".", "=") + "\n\n"); |
---|
| 1127 | buf.append("Warning: No model could be built, hence ZeroR model is used:\n\n"); |
---|
| 1128 | buf.append(m_ZeroR.toString()); |
---|
| 1129 | return buf.toString(); |
---|
| 1130 | } |
---|
| 1131 | |
---|
| 1132 | StringBuffer text = new StringBuffer(); |
---|
| 1133 | |
---|
| 1134 | if (m_NumGenerated == 0) { |
---|
| 1135 | text.append("LogitBoost: No model built yet."); |
---|
| 1136 | // text.append(m_Classifiers[0].toString()+"\n"); |
---|
| 1137 | } else { |
---|
| 1138 | text.append("LogitBoost: Base classifiers and their weights: \n"); |
---|
| 1139 | for (int i = 0; i < m_NumGenerated; i++) { |
---|
| 1140 | text.append("\nIteration "+(i+1)); |
---|
| 1141 | for (int j = 0; j < m_NumClasses; j++) { |
---|
| 1142 | text.append("\n\tClass " + (j + 1) |
---|
| 1143 | + " (" + m_ClassAttribute.name() |
---|
| 1144 | + "=" + m_ClassAttribute.value(j) + ")\n\n" |
---|
| 1145 | + m_Classifiers[j][i].toString() + "\n"); |
---|
| 1146 | } |
---|
| 1147 | } |
---|
| 1148 | text.append("Number of performed iterations: " + |
---|
| 1149 | m_NumGenerated + "\n"); |
---|
| 1150 | } |
---|
| 1151 | |
---|
| 1152 | return text.toString(); |
---|
| 1153 | } |
---|
| 1154 | |
---|
| 1155 | /** |
---|
| 1156 | * Returns the revision string. |
---|
| 1157 | * |
---|
| 1158 | * @return the revision |
---|
| 1159 | */ |
---|
| 1160 | public String getRevision() { |
---|
| 1161 | return RevisionUtils.extract("$Revision: 6091 $"); |
---|
| 1162 | } |
---|
| 1163 | |
---|
| 1164 | /** |
---|
| 1165 | * Main method for testing this class. |
---|
| 1166 | * |
---|
| 1167 | * @param argv the options |
---|
| 1168 | */ |
---|
| 1169 | public static void main(String [] argv) { |
---|
| 1170 | runClassifier(new LogitBoost(), argv); |
---|
| 1171 | } |
---|
| 1172 | } |
---|