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 | * BFTree.java |
---|
19 | * Copyright (C) 2007 Haijian Shi |
---|
20 | * |
---|
21 | */ |
---|
22 | |
---|
23 | package weka.classifiers.trees; |
---|
24 | |
---|
25 | import weka.classifiers.Evaluation; |
---|
26 | import weka.classifiers.RandomizableClassifier; |
---|
27 | import weka.core.AdditionalMeasureProducer; |
---|
28 | import weka.core.Attribute; |
---|
29 | import weka.core.Capabilities; |
---|
30 | import weka.core.FastVector; |
---|
31 | import weka.core.Instance; |
---|
32 | import weka.core.Instances; |
---|
33 | import weka.core.Option; |
---|
34 | import weka.core.RevisionUtils; |
---|
35 | import weka.core.SelectedTag; |
---|
36 | import weka.core.Tag; |
---|
37 | import weka.core.TechnicalInformation; |
---|
38 | import weka.core.TechnicalInformationHandler; |
---|
39 | import weka.core.Utils; |
---|
40 | import weka.core.Capabilities.Capability; |
---|
41 | import weka.core.TechnicalInformation.Field; |
---|
42 | import weka.core.TechnicalInformation.Type; |
---|
43 | import weka.core.matrix.Matrix; |
---|
44 | |
---|
45 | import java.util.Arrays; |
---|
46 | import java.util.Enumeration; |
---|
47 | import java.util.Random; |
---|
48 | import java.util.Vector; |
---|
49 | |
---|
50 | /** |
---|
51 | <!-- globalinfo-start --> |
---|
52 | * Class for building a best-first decision tree classifier. This class uses binary split for both nominal and numeric attributes. For missing values, the method of 'fractional' instances is used.<br/> |
---|
53 | * <br/> |
---|
54 | * For more information, see:<br/> |
---|
55 | * <br/> |
---|
56 | * Haijian Shi (2007). Best-first decision tree learning. Hamilton, NZ.<br/> |
---|
57 | * <br/> |
---|
58 | * Jerome Friedman, Trevor Hastie, Robert Tibshirani (2000). Additive logistic regression : A statistical view of boosting. Annals of statistics. 28(2):337-407. |
---|
59 | * <p/> |
---|
60 | <!-- globalinfo-end --> |
---|
61 | * |
---|
62 | <!-- technical-bibtex-start --> |
---|
63 | * BibTeX: |
---|
64 | * <pre> |
---|
65 | * @mastersthesis{Shi2007, |
---|
66 | * address = {Hamilton, NZ}, |
---|
67 | * author = {Haijian Shi}, |
---|
68 | * note = {COMP594}, |
---|
69 | * school = {University of Waikato}, |
---|
70 | * title = {Best-first decision tree learning}, |
---|
71 | * year = {2007} |
---|
72 | * } |
---|
73 | * |
---|
74 | * @article{Friedman2000, |
---|
75 | * author = {Jerome Friedman and Trevor Hastie and Robert Tibshirani}, |
---|
76 | * journal = {Annals of statistics}, |
---|
77 | * number = {2}, |
---|
78 | * pages = {337-407}, |
---|
79 | * title = {Additive logistic regression : A statistical view of boosting}, |
---|
80 | * volume = {28}, |
---|
81 | * year = {2000}, |
---|
82 | * ISSN = {0090-5364} |
---|
83 | * } |
---|
84 | * </pre> |
---|
85 | * <p/> |
---|
86 | <!-- technical-bibtex-end --> |
---|
87 | * |
---|
88 | <!-- options-start --> |
---|
89 | * Valid options are: <p/> |
---|
90 | * |
---|
91 | * <pre> -S <num> |
---|
92 | * Random number seed. |
---|
93 | * (default 1)</pre> |
---|
94 | * |
---|
95 | * <pre> -D |
---|
96 | * If set, classifier is run in debug mode and |
---|
97 | * may output additional info to the console</pre> |
---|
98 | * |
---|
99 | * <pre> -P <UNPRUNED|POSTPRUNED|PREPRUNED> |
---|
100 | * The pruning strategy. |
---|
101 | * (default: POSTPRUNED)</pre> |
---|
102 | * |
---|
103 | * <pre> -M <min no> |
---|
104 | * The minimal number of instances at the terminal nodes. |
---|
105 | * (default 2)</pre> |
---|
106 | * |
---|
107 | * <pre> -N <num folds> |
---|
108 | * The number of folds used in the pruning. |
---|
109 | * (default 5)</pre> |
---|
110 | * |
---|
111 | * <pre> -H |
---|
112 | * Don't use heuristic search for nominal attributes in multi-class |
---|
113 | * problem (default yes). |
---|
114 | * </pre> |
---|
115 | * |
---|
116 | * <pre> -G |
---|
117 | * Don't use Gini index for splitting (default yes), |
---|
118 | * if not information is used.</pre> |
---|
119 | * |
---|
120 | * <pre> -R |
---|
121 | * Don't use error rate in internal cross-validation (default yes), |
---|
122 | * but root mean squared error.</pre> |
---|
123 | * |
---|
124 | * <pre> -A |
---|
125 | * Use the 1 SE rule to make pruning decision. |
---|
126 | * (default no).</pre> |
---|
127 | * |
---|
128 | * <pre> -C |
---|
129 | * Percentage of training data size (0-1] |
---|
130 | * (default 1).</pre> |
---|
131 | * |
---|
132 | <!-- options-end --> |
---|
133 | * |
---|
134 | * @author Haijian Shi (hs69@cs.waikato.ac.nz) |
---|
135 | * @version $Revision: 5987 $ |
---|
136 | */ |
---|
137 | public class BFTree |
---|
138 | extends RandomizableClassifier |
---|
139 | implements AdditionalMeasureProducer, TechnicalInformationHandler { |
---|
140 | |
---|
141 | /** For serialization. */ |
---|
142 | private static final long serialVersionUID = -7035607375962528217L; |
---|
143 | |
---|
144 | /** pruning strategy: un-pruned */ |
---|
145 | public static final int PRUNING_UNPRUNED = 0; |
---|
146 | /** pruning strategy: post-pruning */ |
---|
147 | public static final int PRUNING_POSTPRUNING = 1; |
---|
148 | /** pruning strategy: pre-pruning */ |
---|
149 | public static final int PRUNING_PREPRUNING = 2; |
---|
150 | /** pruning strategy */ |
---|
151 | public static final Tag[] TAGS_PRUNING = { |
---|
152 | new Tag(PRUNING_UNPRUNED, "unpruned", "Un-pruned"), |
---|
153 | new Tag(PRUNING_POSTPRUNING, "postpruned", "Post-pruning"), |
---|
154 | new Tag(PRUNING_PREPRUNING, "prepruned", "Pre-pruning") |
---|
155 | }; |
---|
156 | |
---|
157 | /** the pruning strategy */ |
---|
158 | protected int m_PruningStrategy = PRUNING_POSTPRUNING; |
---|
159 | |
---|
160 | /** Successor nodes. */ |
---|
161 | protected BFTree[] m_Successors; |
---|
162 | |
---|
163 | /** Attribute used for splitting. */ |
---|
164 | protected Attribute m_Attribute; |
---|
165 | |
---|
166 | /** Split point (for numeric attributes). */ |
---|
167 | protected double m_SplitValue; |
---|
168 | |
---|
169 | /** Split subset (for nominal attributes). */ |
---|
170 | protected String m_SplitString; |
---|
171 | |
---|
172 | /** Class value for a node. */ |
---|
173 | protected double m_ClassValue; |
---|
174 | |
---|
175 | /** Class attribute of a dataset. */ |
---|
176 | protected Attribute m_ClassAttribute; |
---|
177 | |
---|
178 | /** Minimum number of instances at leaf nodes. */ |
---|
179 | protected int m_minNumObj = 2; |
---|
180 | |
---|
181 | /** Number of folds for the pruning. */ |
---|
182 | protected int m_numFoldsPruning = 5; |
---|
183 | |
---|
184 | /** If the ndoe is leaf node. */ |
---|
185 | protected boolean m_isLeaf; |
---|
186 | |
---|
187 | /** Number of expansions. */ |
---|
188 | protected static int m_Expansion; |
---|
189 | |
---|
190 | /** Fixed number of expansions (if no pruning method is used, its value is -1. Otherwise, |
---|
191 | * its value is gotten from internal cross-validation). */ |
---|
192 | protected int m_FixedExpansion = -1; |
---|
193 | |
---|
194 | /** If use huristic search for binary split (default true). Note even if its value is true, it is only |
---|
195 | * used when the number of values of a nominal attribute is larger than 4. */ |
---|
196 | protected boolean m_Heuristic = true; |
---|
197 | |
---|
198 | /** If use Gini index as the splitting criterion - default (if not, information is used). */ |
---|
199 | protected boolean m_UseGini = true; |
---|
200 | |
---|
201 | /** If use error rate in internal cross-validation to fix the number of expansions - default |
---|
202 | * (if not, root mean squared error is used). */ |
---|
203 | protected boolean m_UseErrorRate = true; |
---|
204 | |
---|
205 | /** If use the 1SE rule to make the decision. */ |
---|
206 | protected boolean m_UseOneSE = false; |
---|
207 | |
---|
208 | /** Class distributions. */ |
---|
209 | protected double[] m_Distribution; |
---|
210 | |
---|
211 | /** Branch proportions. */ |
---|
212 | protected double[] m_Props; |
---|
213 | |
---|
214 | /** Sorted indices. */ |
---|
215 | protected int[][] m_SortedIndices; |
---|
216 | |
---|
217 | /** Sorted weights. */ |
---|
218 | protected double[][] m_Weights; |
---|
219 | |
---|
220 | /** Distributions of each attribute for two successor nodes. */ |
---|
221 | protected double[][][] m_Dists; |
---|
222 | |
---|
223 | /** Class probabilities. */ |
---|
224 | protected double[] m_ClassProbs; |
---|
225 | |
---|
226 | /** Total weights. */ |
---|
227 | protected double m_TotalWeight; |
---|
228 | |
---|
229 | /** The training data size (0-1). Default 1. */ |
---|
230 | protected double m_SizePer = 1; |
---|
231 | |
---|
232 | /** |
---|
233 | * Returns a string describing classifier |
---|
234 | * |
---|
235 | * @return a description suitable for displaying in the |
---|
236 | * explorer/experimenter gui |
---|
237 | */ |
---|
238 | public String globalInfo() { |
---|
239 | return |
---|
240 | "Class for building a best-first decision tree classifier. " |
---|
241 | + "This class uses binary split for both nominal and numeric attributes. " |
---|
242 | + "For missing values, the method of 'fractional' instances is used.\n\n" |
---|
243 | + "For more information, see:\n\n" |
---|
244 | + getTechnicalInformation().toString(); |
---|
245 | } |
---|
246 | |
---|
247 | /** |
---|
248 | * Returns an instance of a TechnicalInformation object, containing |
---|
249 | * detailed information about the technical background of this class, |
---|
250 | * e.g., paper reference or book this class is based on. |
---|
251 | * |
---|
252 | * @return the technical information about this class |
---|
253 | */ |
---|
254 | public TechnicalInformation getTechnicalInformation() { |
---|
255 | TechnicalInformation result; |
---|
256 | TechnicalInformation additional; |
---|
257 | |
---|
258 | result = new TechnicalInformation(Type.MASTERSTHESIS); |
---|
259 | result.setValue(Field.AUTHOR, "Haijian Shi"); |
---|
260 | result.setValue(Field.YEAR, "2007"); |
---|
261 | result.setValue(Field.TITLE, "Best-first decision tree learning"); |
---|
262 | result.setValue(Field.SCHOOL, "University of Waikato"); |
---|
263 | result.setValue(Field.ADDRESS, "Hamilton, NZ"); |
---|
264 | result.setValue(Field.NOTE, "COMP594"); |
---|
265 | |
---|
266 | additional = result.add(Type.ARTICLE); |
---|
267 | additional.setValue(Field.AUTHOR, "Jerome Friedman and Trevor Hastie and Robert Tibshirani"); |
---|
268 | additional.setValue(Field.YEAR, "2000"); |
---|
269 | additional.setValue(Field.TITLE, "Additive logistic regression : A statistical view of boosting"); |
---|
270 | additional.setValue(Field.JOURNAL, "Annals of statistics"); |
---|
271 | additional.setValue(Field.VOLUME, "28"); |
---|
272 | additional.setValue(Field.NUMBER, "2"); |
---|
273 | additional.setValue(Field.PAGES, "337-407"); |
---|
274 | additional.setValue(Field.ISSN, "0090-5364"); |
---|
275 | |
---|
276 | return result; |
---|
277 | } |
---|
278 | |
---|
279 | /** |
---|
280 | * Returns default capabilities of the classifier. |
---|
281 | * |
---|
282 | * @return the capabilities of this classifier |
---|
283 | */ |
---|
284 | public Capabilities getCapabilities() { |
---|
285 | Capabilities result = super.getCapabilities(); |
---|
286 | result.disableAll(); |
---|
287 | |
---|
288 | // attributes |
---|
289 | result.enable(Capability.NOMINAL_ATTRIBUTES); |
---|
290 | result.enable(Capability.NUMERIC_ATTRIBUTES); |
---|
291 | result.enable(Capability.MISSING_VALUES); |
---|
292 | |
---|
293 | // class |
---|
294 | result.enable(Capability.NOMINAL_CLASS); |
---|
295 | |
---|
296 | return result; |
---|
297 | } |
---|
298 | |
---|
299 | /** |
---|
300 | * Method for building a BestFirst decision tree classifier. |
---|
301 | * |
---|
302 | * @param data set of instances serving as training data |
---|
303 | * @throws Exception if decision tree cannot be built successfully |
---|
304 | */ |
---|
305 | public void buildClassifier(Instances data) throws Exception { |
---|
306 | |
---|
307 | getCapabilities().testWithFail(data); |
---|
308 | data = new Instances(data); |
---|
309 | data.deleteWithMissingClass(); |
---|
310 | |
---|
311 | // build an unpruned tree |
---|
312 | if (m_PruningStrategy == PRUNING_UNPRUNED) { |
---|
313 | |
---|
314 | // calculate sorted indices, weights and initial class probabilities |
---|
315 | int[][] sortedIndices = new int[data.numAttributes()][0]; |
---|
316 | double[][] weights = new double[data.numAttributes()][0]; |
---|
317 | double[] classProbs = new double[data.numClasses()]; |
---|
318 | double totalWeight = computeSortedInfo(data,sortedIndices, weights,classProbs); |
---|
319 | |
---|
320 | // Compute information of the best split for this node (include split attribute, |
---|
321 | // split value and gini gain (or information gain)). At the same time, compute |
---|
322 | // variables dists, props and totalSubsetWeights. |
---|
323 | double[][][] dists = new double[data.numAttributes()][2][data.numClasses()]; |
---|
324 | double[][] props = new double[data.numAttributes()][2]; |
---|
325 | double[][] totalSubsetWeights = new double[data.numAttributes()][2]; |
---|
326 | FastVector nodeInfo = computeSplitInfo(this, data, sortedIndices, weights, dists, |
---|
327 | props, totalSubsetWeights, m_Heuristic, m_UseGini); |
---|
328 | |
---|
329 | // add the node (with all split info) into BestFirstElements |
---|
330 | FastVector BestFirstElements = new FastVector(); |
---|
331 | BestFirstElements.addElement(nodeInfo); |
---|
332 | |
---|
333 | // Make the best-first decision tree. |
---|
334 | int attIndex = ((Attribute)nodeInfo.elementAt(1)).index(); |
---|
335 | m_Expansion = 0; |
---|
336 | makeTree(BestFirstElements, data, sortedIndices, weights, dists, classProbs, |
---|
337 | totalWeight, props[attIndex] ,m_minNumObj, m_Heuristic, m_UseGini, m_FixedExpansion); |
---|
338 | |
---|
339 | return; |
---|
340 | } |
---|
341 | |
---|
342 | // the following code is for pre-pruning and post-pruning methods |
---|
343 | |
---|
344 | // Compute train data, test data, sorted indices, sorted weights, total weights, |
---|
345 | // class probabilities, class distributions, branch proportions and total subset |
---|
346 | // weights for root nodes of each fold for prepruning and postpruning. |
---|
347 | int expansion = 0; |
---|
348 | |
---|
349 | Random random = new Random(m_Seed); |
---|
350 | Instances cvData = new Instances(data); |
---|
351 | cvData.randomize(random); |
---|
352 | cvData = new Instances(cvData,0,(int)(cvData.numInstances()*m_SizePer)-1); |
---|
353 | cvData.stratify(m_numFoldsPruning); |
---|
354 | |
---|
355 | Instances[] train = new Instances[m_numFoldsPruning]; |
---|
356 | Instances[] test = new Instances[m_numFoldsPruning]; |
---|
357 | FastVector[] parallelBFElements = new FastVector [m_numFoldsPruning]; |
---|
358 | BFTree[] m_roots = new BFTree[m_numFoldsPruning]; |
---|
359 | |
---|
360 | int[][][] sortedIndices = new int[m_numFoldsPruning][data.numAttributes()][0]; |
---|
361 | double[][][] weights = new double[m_numFoldsPruning][data.numAttributes()][0]; |
---|
362 | double[][] classProbs = new double[m_numFoldsPruning][data.numClasses()]; |
---|
363 | double[] totalWeight = new double[m_numFoldsPruning]; |
---|
364 | |
---|
365 | double[][][][] dists = |
---|
366 | new double[m_numFoldsPruning][data.numAttributes()][2][data.numClasses()]; |
---|
367 | double[][][] props = |
---|
368 | new double[m_numFoldsPruning][data.numAttributes()][2]; |
---|
369 | double[][][] totalSubsetWeights = |
---|
370 | new double[m_numFoldsPruning][data.numAttributes()][2]; |
---|
371 | FastVector[] nodeInfo = new FastVector[m_numFoldsPruning]; |
---|
372 | |
---|
373 | for (int i = 0; i < m_numFoldsPruning; i++) { |
---|
374 | train[i] = cvData.trainCV(m_numFoldsPruning, i); |
---|
375 | test[i] = cvData.testCV(m_numFoldsPruning, i); |
---|
376 | parallelBFElements[i] = new FastVector(); |
---|
377 | m_roots[i] = new BFTree(); |
---|
378 | |
---|
379 | // calculate sorted indices, weights, initial class counts and total weights for each training data |
---|
380 | totalWeight[i] = computeSortedInfo(train[i],sortedIndices[i], weights[i], |
---|
381 | classProbs[i]); |
---|
382 | |
---|
383 | // compute information of the best split for this node (include split attribute, |
---|
384 | // split value and gini gain (or information gain)) in this fold |
---|
385 | nodeInfo[i] = computeSplitInfo(m_roots[i], train[i], sortedIndices[i], |
---|
386 | weights[i], dists[i], props[i], totalSubsetWeights[i], m_Heuristic, m_UseGini); |
---|
387 | |
---|
388 | // compute information for root nodes |
---|
389 | |
---|
390 | int attIndex = ((Attribute)nodeInfo[i].elementAt(1)).index(); |
---|
391 | |
---|
392 | m_roots[i].m_SortedIndices = new int[sortedIndices[i].length][0]; |
---|
393 | m_roots[i].m_Weights = new double[weights[i].length][0]; |
---|
394 | m_roots[i].m_Dists = new double[dists[i].length][0][0]; |
---|
395 | m_roots[i].m_ClassProbs = new double[classProbs[i].length]; |
---|
396 | m_roots[i].m_Distribution = new double[classProbs[i].length]; |
---|
397 | m_roots[i].m_Props = new double[2]; |
---|
398 | |
---|
399 | for (int j=0; j<m_roots[i].m_SortedIndices.length; j++) { |
---|
400 | m_roots[i].m_SortedIndices[j] = sortedIndices[i][j]; |
---|
401 | m_roots[i].m_Weights[j] = weights[i][j]; |
---|
402 | m_roots[i].m_Dists[j] = dists[i][j]; |
---|
403 | } |
---|
404 | |
---|
405 | System.arraycopy(classProbs[i], 0, m_roots[i].m_ClassProbs, 0, |
---|
406 | classProbs[i].length); |
---|
407 | if (Utils.sum(m_roots[i].m_ClassProbs)!=0) |
---|
408 | Utils.normalize(m_roots[i].m_ClassProbs); |
---|
409 | |
---|
410 | System.arraycopy(classProbs[i], 0, m_roots[i].m_Distribution, 0, |
---|
411 | classProbs[i].length); |
---|
412 | System.arraycopy(props[i][attIndex], 0, m_roots[i].m_Props, 0, |
---|
413 | props[i][attIndex].length); |
---|
414 | |
---|
415 | m_roots[i].m_TotalWeight = totalWeight[i]; |
---|
416 | |
---|
417 | parallelBFElements[i].addElement(nodeInfo[i]); |
---|
418 | } |
---|
419 | |
---|
420 | // build a pre-pruned tree |
---|
421 | if (m_PruningStrategy == PRUNING_PREPRUNING) { |
---|
422 | |
---|
423 | double previousError = Double.MAX_VALUE; |
---|
424 | double currentError = previousError; |
---|
425 | double minError = Double.MAX_VALUE; |
---|
426 | int minExpansion = 0; |
---|
427 | FastVector errorList = new FastVector(); |
---|
428 | while(true) { |
---|
429 | // compute average error |
---|
430 | double expansionError = 0; |
---|
431 | int count = 0; |
---|
432 | |
---|
433 | for (int i=0; i<m_numFoldsPruning; i++) { |
---|
434 | Evaluation eval; |
---|
435 | |
---|
436 | // calculate error rate if only root node |
---|
437 | if (expansion==0) { |
---|
438 | m_roots[i].m_isLeaf = true; |
---|
439 | eval = new Evaluation(test[i]); |
---|
440 | eval.evaluateModel(m_roots[i], test[i]); |
---|
441 | if (m_UseErrorRate) expansionError += eval.errorRate(); |
---|
442 | else expansionError += eval.rootMeanSquaredError(); |
---|
443 | count ++; |
---|
444 | } |
---|
445 | |
---|
446 | // make tree - expand one node at a time |
---|
447 | else { |
---|
448 | if (m_roots[i] == null) continue; // if the tree cannot be expanded, go to next fold |
---|
449 | m_roots[i].m_isLeaf = false; |
---|
450 | BFTree nodeToSplit = (BFTree) |
---|
451 | (((FastVector)(parallelBFElements[i].elementAt(0))).elementAt(0)); |
---|
452 | if (!m_roots[i].makeTree(parallelBFElements[i], m_roots[i], train[i], |
---|
453 | nodeToSplit.m_SortedIndices, nodeToSplit.m_Weights, |
---|
454 | nodeToSplit.m_Dists, nodeToSplit.m_ClassProbs, |
---|
455 | nodeToSplit.m_TotalWeight, nodeToSplit.m_Props, m_minNumObj, |
---|
456 | m_Heuristic, m_UseGini)) { |
---|
457 | m_roots[i] = null; // cannot be expanded |
---|
458 | continue; |
---|
459 | } |
---|
460 | eval = new Evaluation(test[i]); |
---|
461 | eval.evaluateModel(m_roots[i], test[i]); |
---|
462 | if (m_UseErrorRate) expansionError += eval.errorRate(); |
---|
463 | else expansionError += eval.rootMeanSquaredError(); |
---|
464 | count ++; |
---|
465 | } |
---|
466 | } |
---|
467 | |
---|
468 | // no tree can be expanded any more |
---|
469 | if (count==0) break; |
---|
470 | |
---|
471 | expansionError /=count; |
---|
472 | errorList.addElement(new Double(expansionError)); |
---|
473 | currentError = expansionError; |
---|
474 | |
---|
475 | if (!m_UseOneSE) { |
---|
476 | if (currentError>previousError) |
---|
477 | break; |
---|
478 | } |
---|
479 | |
---|
480 | else { |
---|
481 | if (expansionError < minError) { |
---|
482 | minError = expansionError; |
---|
483 | minExpansion = expansion; |
---|
484 | } |
---|
485 | |
---|
486 | if (currentError>previousError) { |
---|
487 | double oneSE = Math.sqrt(minError*(1-minError)/ |
---|
488 | data.numInstances()); |
---|
489 | if (currentError > minError + oneSE) { |
---|
490 | break; |
---|
491 | } |
---|
492 | } |
---|
493 | } |
---|
494 | |
---|
495 | expansion ++; |
---|
496 | previousError = currentError; |
---|
497 | } |
---|
498 | |
---|
499 | if (!m_UseOneSE) expansion = expansion - 1; |
---|
500 | else { |
---|
501 | double oneSE = Math.sqrt(minError*(1-minError)/data.numInstances()); |
---|
502 | for (int i=0; i<errorList.size(); i++) { |
---|
503 | double error = ((Double)(errorList.elementAt(i))).doubleValue(); |
---|
504 | if (error<=minError + oneSE) { // && counts[i]>=m_numFoldsPruning/2) { |
---|
505 | expansion = i; |
---|
506 | break; |
---|
507 | } |
---|
508 | } |
---|
509 | } |
---|
510 | } |
---|
511 | |
---|
512 | // build a postpruned tree |
---|
513 | else { |
---|
514 | FastVector[] modelError = new FastVector[m_numFoldsPruning]; |
---|
515 | |
---|
516 | // calculate error of each expansion for each fold |
---|
517 | for (int i = 0; i < m_numFoldsPruning; i++) { |
---|
518 | modelError[i] = new FastVector(); |
---|
519 | |
---|
520 | m_roots[i].m_isLeaf = true; |
---|
521 | Evaluation eval = new Evaluation(test[i]); |
---|
522 | eval.evaluateModel(m_roots[i], test[i]); |
---|
523 | double error; |
---|
524 | if (m_UseErrorRate) error = eval.errorRate(); |
---|
525 | else error = eval.rootMeanSquaredError(); |
---|
526 | modelError[i].addElement(new Double(error)); |
---|
527 | |
---|
528 | m_roots[i].m_isLeaf = false; |
---|
529 | BFTree nodeToSplit = (BFTree) |
---|
530 | (((FastVector)(parallelBFElements[i].elementAt(0))).elementAt(0)); |
---|
531 | |
---|
532 | m_roots[i].makeTree(parallelBFElements[i], m_roots[i], train[i], test[i], |
---|
533 | modelError[i],nodeToSplit.m_SortedIndices, nodeToSplit.m_Weights, |
---|
534 | nodeToSplit.m_Dists, nodeToSplit.m_ClassProbs, |
---|
535 | nodeToSplit.m_TotalWeight, nodeToSplit.m_Props, m_minNumObj, |
---|
536 | m_Heuristic, m_UseGini, m_UseErrorRate); |
---|
537 | m_roots[i] = null; |
---|
538 | } |
---|
539 | |
---|
540 | // find the expansion with minimal error rate |
---|
541 | double minError = Double.MAX_VALUE; |
---|
542 | |
---|
543 | int maxExpansion = modelError[0].size(); |
---|
544 | for (int i=1; i<modelError.length; i++) { |
---|
545 | if (modelError[i].size()>maxExpansion) |
---|
546 | maxExpansion = modelError[i].size(); |
---|
547 | } |
---|
548 | |
---|
549 | double[] error = new double[maxExpansion]; |
---|
550 | int[] counts = new int[maxExpansion]; |
---|
551 | for (int i=0; i<maxExpansion; i++) { |
---|
552 | counts[i] = 0; |
---|
553 | error[i] = 0; |
---|
554 | for (int j=0; j<m_numFoldsPruning; j++) { |
---|
555 | if (i<modelError[j].size()) { |
---|
556 | error[i] += ((Double)modelError[j].elementAt(i)).doubleValue(); |
---|
557 | counts[i]++; |
---|
558 | } |
---|
559 | } |
---|
560 | error[i] = error[i]/counts[i]; //average error for each expansion |
---|
561 | |
---|
562 | if (error[i]<minError) {// && counts[i]>=m_numFoldsPruning/2) { |
---|
563 | minError = error[i]; |
---|
564 | expansion = i; |
---|
565 | } |
---|
566 | } |
---|
567 | |
---|
568 | // the 1 SE rule choosen |
---|
569 | if (m_UseOneSE) { |
---|
570 | double oneSE = Math.sqrt(minError*(1-minError)/ |
---|
571 | data.numInstances()); |
---|
572 | for (int i=0; i<maxExpansion; i++) { |
---|
573 | if (error[i]<=minError + oneSE) { // && counts[i]>=m_numFoldsPruning/2) { |
---|
574 | expansion = i; |
---|
575 | break; |
---|
576 | } |
---|
577 | } |
---|
578 | } |
---|
579 | } |
---|
580 | |
---|
581 | // make tree on all data based on the expansion caculated |
---|
582 | // from cross-validation |
---|
583 | |
---|
584 | // calculate sorted indices, weights and initial class counts |
---|
585 | int[][] prune_sortedIndices = new int[data.numAttributes()][0]; |
---|
586 | double[][] prune_weights = new double[data.numAttributes()][0]; |
---|
587 | double[] prune_classProbs = new double[data.numClasses()]; |
---|
588 | double prune_totalWeight = computeSortedInfo(data, prune_sortedIndices, |
---|
589 | prune_weights, prune_classProbs); |
---|
590 | |
---|
591 | // compute information of the best split for this node (include split attribute, |
---|
592 | // split value and gini gain) |
---|
593 | double[][][] prune_dists = new double[data.numAttributes()][2][data.numClasses()]; |
---|
594 | double[][] prune_props = new double[data.numAttributes()][2]; |
---|
595 | double[][] prune_totalSubsetWeights = new double[data.numAttributes()][2]; |
---|
596 | FastVector prune_nodeInfo = computeSplitInfo(this, data, prune_sortedIndices, |
---|
597 | prune_weights, prune_dists, prune_props, prune_totalSubsetWeights, m_Heuristic,m_UseGini); |
---|
598 | |
---|
599 | // add the root node (with its split info) to BestFirstElements |
---|
600 | FastVector BestFirstElements = new FastVector(); |
---|
601 | BestFirstElements.addElement(prune_nodeInfo); |
---|
602 | |
---|
603 | int attIndex = ((Attribute)prune_nodeInfo.elementAt(1)).index(); |
---|
604 | m_Expansion = 0; |
---|
605 | makeTree(BestFirstElements, data, prune_sortedIndices, prune_weights, prune_dists, |
---|
606 | prune_classProbs, prune_totalWeight, prune_props[attIndex] ,m_minNumObj, |
---|
607 | m_Heuristic, m_UseGini, expansion); |
---|
608 | } |
---|
609 | |
---|
610 | /** |
---|
611 | * Recursively build a best-first decision tree. |
---|
612 | * Method for building a Best-First tree for a given number of expansions. |
---|
613 | * preExpasion is -1 means that no expansion is specified (just for a |
---|
614 | * tree without any pruning method). Pre-pruning and post-pruning methods also |
---|
615 | * use this method to build the final tree on all training data based on the |
---|
616 | * expansion calculated from internal cross-validation. |
---|
617 | * |
---|
618 | * @param BestFirstElements list to store BFTree nodes |
---|
619 | * @param data training data |
---|
620 | * @param sortedIndices sorted indices of the instances |
---|
621 | * @param weights weights of the instances |
---|
622 | * @param dists class distributions for each attribute |
---|
623 | * @param classProbs class probabilities of this node |
---|
624 | * @param totalWeight total weight of this node (note if the node |
---|
625 | * can not split, this value is not calculated.) |
---|
626 | * @param branchProps proportions of two subbranches |
---|
627 | * @param minNumObj minimal number of instances at leaf nodes |
---|
628 | * @param useHeuristic if use heuristic search for nominal attributes |
---|
629 | * in multi-class problem |
---|
630 | * @param useGini if use Gini index as splitting criterion |
---|
631 | * @param preExpansion the number of expansions the tree to be expanded |
---|
632 | * @throws Exception if something goes wrong |
---|
633 | */ |
---|
634 | protected void makeTree(FastVector BestFirstElements,Instances data, |
---|
635 | int[][] sortedIndices, double[][] weights, double[][][] dists, |
---|
636 | double[] classProbs, double totalWeight, double[] branchProps, |
---|
637 | int minNumObj, boolean useHeuristic, boolean useGini, int preExpansion) |
---|
638 | throws Exception { |
---|
639 | |
---|
640 | if (BestFirstElements.size()==0) return; |
---|
641 | |
---|
642 | /////////////////////////////////////////////////////////////////////// |
---|
643 | // All information about the node to split (the first BestFirst object in |
---|
644 | // BestFirstElements) |
---|
645 | FastVector firstElement = (FastVector)BestFirstElements.elementAt(0); |
---|
646 | |
---|
647 | // split attribute |
---|
648 | Attribute att = (Attribute)firstElement.elementAt(1); |
---|
649 | |
---|
650 | // info of split value or split string |
---|
651 | double splitValue = Double.NaN; |
---|
652 | String splitStr = null; |
---|
653 | if (att.isNumeric()) |
---|
654 | splitValue = ((Double)firstElement.elementAt(2)).doubleValue(); |
---|
655 | else { |
---|
656 | splitStr=((String)firstElement.elementAt(2)).toString(); |
---|
657 | } |
---|
658 | |
---|
659 | // the best gini gain or information gain of this node |
---|
660 | double gain = ((Double)firstElement.elementAt(3)).doubleValue(); |
---|
661 | /////////////////////////////////////////////////////////////////////// |
---|
662 | |
---|
663 | if (m_ClassProbs==null) { |
---|
664 | m_SortedIndices = new int[sortedIndices.length][0]; |
---|
665 | m_Weights = new double[weights.length][0]; |
---|
666 | m_Dists = new double[dists.length][0][0]; |
---|
667 | m_ClassProbs = new double[classProbs.length]; |
---|
668 | m_Distribution = new double[classProbs.length]; |
---|
669 | m_Props = new double[2]; |
---|
670 | |
---|
671 | for (int i=0; i<m_SortedIndices.length; i++) { |
---|
672 | m_SortedIndices[i] = sortedIndices[i]; |
---|
673 | m_Weights[i] = weights[i]; |
---|
674 | m_Dists[i] = dists[i]; |
---|
675 | } |
---|
676 | |
---|
677 | System.arraycopy(classProbs, 0, m_ClassProbs, 0, classProbs.length); |
---|
678 | System.arraycopy(classProbs, 0, m_Distribution, 0, classProbs.length); |
---|
679 | System.arraycopy(branchProps, 0, m_Props, 0, m_Props.length); |
---|
680 | m_TotalWeight = totalWeight; |
---|
681 | if (Utils.sum(m_ClassProbs)!=0) Utils.normalize(m_ClassProbs); |
---|
682 | } |
---|
683 | |
---|
684 | // If no enough data or this node can not be split, find next node to split. |
---|
685 | if (totalWeight < 2*minNumObj || branchProps[0]==0 |
---|
686 | || branchProps[1]==0) { |
---|
687 | // remove the first element |
---|
688 | BestFirstElements.removeElementAt(0); |
---|
689 | |
---|
690 | makeLeaf(data); |
---|
691 | if (BestFirstElements.size()!=0) { |
---|
692 | FastVector nextSplitElement = (FastVector)BestFirstElements.elementAt(0); |
---|
693 | BFTree nextSplitNode = (BFTree)nextSplitElement.elementAt(0); |
---|
694 | nextSplitNode.makeTree(BestFirstElements,data, |
---|
695 | nextSplitNode.m_SortedIndices, nextSplitNode.m_Weights, |
---|
696 | nextSplitNode.m_Dists, |
---|
697 | nextSplitNode.m_ClassProbs, nextSplitNode.m_TotalWeight, |
---|
698 | nextSplitNode.m_Props, minNumObj, useHeuristic, useGini, preExpansion); |
---|
699 | } |
---|
700 | return; |
---|
701 | } |
---|
702 | |
---|
703 | // If gini gain or information gain is 0, make all nodes in the BestFirstElements leaf nodes |
---|
704 | // because these nodes are sorted descendingly according to gini gain or information gain. |
---|
705 | // (namely, gini gain or information gain of all nodes in BestFirstEelements is 0). |
---|
706 | if (gain==0 || preExpansion==m_Expansion) { |
---|
707 | for (int i=0; i<BestFirstElements.size(); i++) { |
---|
708 | FastVector element = (FastVector)BestFirstElements.elementAt(i); |
---|
709 | BFTree node = (BFTree)element.elementAt(0); |
---|
710 | node.makeLeaf(data); |
---|
711 | } |
---|
712 | BestFirstElements.removeAllElements(); |
---|
713 | } |
---|
714 | |
---|
715 | // gain is not 0 |
---|
716 | else { |
---|
717 | // remove the first element |
---|
718 | BestFirstElements.removeElementAt(0); |
---|
719 | |
---|
720 | m_Attribute = att; |
---|
721 | if (m_Attribute.isNumeric()) m_SplitValue = splitValue; |
---|
722 | else m_SplitString = splitStr; |
---|
723 | |
---|
724 | int[][][] subsetIndices = new int[2][data.numAttributes()][0]; |
---|
725 | double[][][] subsetWeights = new double[2][data.numAttributes()][0]; |
---|
726 | |
---|
727 | splitData(subsetIndices, subsetWeights, m_Attribute, m_SplitValue, |
---|
728 | m_SplitString, sortedIndices, weights, data); |
---|
729 | |
---|
730 | // If split will generate node(s) which has total weights less than m_minNumObj, |
---|
731 | // do not split. |
---|
732 | int attIndex = att.index(); |
---|
733 | if (subsetIndices[0][attIndex].length<minNumObj || |
---|
734 | subsetIndices[1][attIndex].length<minNumObj) { |
---|
735 | makeLeaf(data); |
---|
736 | } |
---|
737 | |
---|
738 | // split the node |
---|
739 | else { |
---|
740 | m_isLeaf = false; |
---|
741 | m_Attribute = att; |
---|
742 | |
---|
743 | // if expansion is specified (if pruning method used) |
---|
744 | if ( (m_PruningStrategy == PRUNING_PREPRUNING) |
---|
745 | || (m_PruningStrategy == PRUNING_POSTPRUNING) |
---|
746 | || (preExpansion != -1)) |
---|
747 | m_Expansion++; |
---|
748 | |
---|
749 | makeSuccessors(BestFirstElements,data,subsetIndices,subsetWeights,dists, |
---|
750 | att,useHeuristic, useGini); |
---|
751 | } |
---|
752 | |
---|
753 | // choose next node to split |
---|
754 | if (BestFirstElements.size()!=0) { |
---|
755 | FastVector nextSplitElement = (FastVector)BestFirstElements.elementAt(0); |
---|
756 | BFTree nextSplitNode = (BFTree)nextSplitElement.elementAt(0); |
---|
757 | nextSplitNode.makeTree(BestFirstElements,data, |
---|
758 | nextSplitNode.m_SortedIndices, nextSplitNode.m_Weights, |
---|
759 | nextSplitNode.m_Dists, |
---|
760 | nextSplitNode.m_ClassProbs, nextSplitNode.m_TotalWeight, |
---|
761 | nextSplitNode.m_Props, minNumObj, useHeuristic, useGini, preExpansion); |
---|
762 | } |
---|
763 | |
---|
764 | } |
---|
765 | } |
---|
766 | |
---|
767 | /** |
---|
768 | * This method is to find the number of expansions based on internal |
---|
769 | * cross-validation for just pre-pruning. It expands the first BestFirst |
---|
770 | * node in the BestFirstElements if it is expansible, otherwise it looks |
---|
771 | * for next exapansible node. If it finds a node is expansibel, expand the |
---|
772 | * node, then return true. (note it just expands one node at a time). |
---|
773 | * |
---|
774 | * @param BestFirstElements list to store BFTree nodes |
---|
775 | * @param root root node of tree in each fold |
---|
776 | * @param train training data |
---|
777 | * @param sortedIndices sorted indices of the instances |
---|
778 | * @param weights weights of the instances |
---|
779 | * @param dists class distributions for each attribute |
---|
780 | * @param classProbs class probabilities of this node |
---|
781 | * @param totalWeight total weight of this node (note if the node |
---|
782 | * can not split, this value is not calculated.) |
---|
783 | * @param branchProps proportions of two subbranches |
---|
784 | * @param minNumObj minimal number of instances at leaf nodes |
---|
785 | * @param useHeuristic if use heuristic search for nominal attributes |
---|
786 | * in multi-class problem |
---|
787 | * @param useGini if use Gini index as splitting criterion |
---|
788 | * @return true if expand successfully, otherwise return false |
---|
789 | * (all nodes in BestFirstElements cannot be |
---|
790 | * expanded). |
---|
791 | * @throws Exception if something goes wrong |
---|
792 | */ |
---|
793 | protected boolean makeTree(FastVector BestFirstElements, BFTree root, |
---|
794 | Instances train, int[][] sortedIndices, double[][] weights, |
---|
795 | double[][][] dists, double[] classProbs, double totalWeight, |
---|
796 | double[] branchProps, int minNumObj, boolean useHeuristic, boolean useGini) |
---|
797 | throws Exception { |
---|
798 | |
---|
799 | if (BestFirstElements.size()==0) return false; |
---|
800 | |
---|
801 | /////////////////////////////////////////////////////////////////////// |
---|
802 | // All information about the node to split (first BestFirst object in |
---|
803 | // BestFirstElements) |
---|
804 | FastVector firstElement = (FastVector)BestFirstElements.elementAt(0); |
---|
805 | |
---|
806 | // node to split |
---|
807 | BFTree nodeToSplit = (BFTree)firstElement.elementAt(0); |
---|
808 | |
---|
809 | // split attribute |
---|
810 | Attribute att = (Attribute)firstElement.elementAt(1); |
---|
811 | |
---|
812 | // info of split value or split string |
---|
813 | double splitValue = Double.NaN; |
---|
814 | String splitStr = null; |
---|
815 | if (att.isNumeric()) |
---|
816 | splitValue = ((Double)firstElement.elementAt(2)).doubleValue(); |
---|
817 | else { |
---|
818 | splitStr=((String)firstElement.elementAt(2)).toString(); |
---|
819 | } |
---|
820 | |
---|
821 | // the best gini gain or information gain of this node |
---|
822 | double gain = ((Double)firstElement.elementAt(3)).doubleValue(); |
---|
823 | /////////////////////////////////////////////////////////////////////// |
---|
824 | |
---|
825 | // If no enough data to split for this node or this node can not be split find next node to split. |
---|
826 | if (totalWeight < 2*minNumObj || branchProps[0]==0 |
---|
827 | || branchProps[1]==0) { |
---|
828 | // remove the first element |
---|
829 | BestFirstElements.removeElementAt(0); |
---|
830 | nodeToSplit.makeLeaf(train); |
---|
831 | BFTree nextNode = (BFTree) |
---|
832 | ((FastVector)BestFirstElements.elementAt(0)).elementAt(0); |
---|
833 | return root.makeTree(BestFirstElements, root, train, |
---|
834 | nextNode.m_SortedIndices, nextNode.m_Weights, nextNode.m_Dists, |
---|
835 | nextNode.m_ClassProbs, nextNode.m_TotalWeight, |
---|
836 | nextNode.m_Props, minNumObj, useHeuristic, useGini); |
---|
837 | } |
---|
838 | |
---|
839 | // If gini gain or information is 0, make all nodes in the BestFirstElements leaf nodes |
---|
840 | // because these node sorted descendingly according to gini gain or information gain. |
---|
841 | // (namely, gini gain or information gain of all nodes in BestFirstEelements is 0). |
---|
842 | if (gain==0) { |
---|
843 | for (int i=0; i<BestFirstElements.size(); i++) { |
---|
844 | FastVector element = (FastVector)BestFirstElements.elementAt(i); |
---|
845 | BFTree node = (BFTree)element.elementAt(0); |
---|
846 | node.makeLeaf(train); |
---|
847 | } |
---|
848 | BestFirstElements.removeAllElements(); |
---|
849 | return false; |
---|
850 | } |
---|
851 | |
---|
852 | else { |
---|
853 | // remove the first element |
---|
854 | BestFirstElements.removeElementAt(0); |
---|
855 | nodeToSplit.m_Attribute = att; |
---|
856 | if (att.isNumeric()) nodeToSplit.m_SplitValue = splitValue; |
---|
857 | else nodeToSplit.m_SplitString = splitStr; |
---|
858 | |
---|
859 | int[][][] subsetIndices = new int[2][train.numAttributes()][0]; |
---|
860 | double[][][] subsetWeights = new double[2][train.numAttributes()][0]; |
---|
861 | |
---|
862 | splitData(subsetIndices, subsetWeights, nodeToSplit.m_Attribute, |
---|
863 | nodeToSplit.m_SplitValue, nodeToSplit.m_SplitString, |
---|
864 | nodeToSplit.m_SortedIndices, nodeToSplit.m_Weights, train); |
---|
865 | |
---|
866 | // if split will generate node(s) which has total weights less than m_minNumObj, |
---|
867 | // do not split |
---|
868 | int attIndex = att.index(); |
---|
869 | if (subsetIndices[0][attIndex].length<minNumObj || |
---|
870 | subsetIndices[1][attIndex].length<minNumObj) { |
---|
871 | |
---|
872 | nodeToSplit.makeLeaf(train); |
---|
873 | BFTree nextNode = (BFTree) |
---|
874 | ((FastVector)BestFirstElements.elementAt(0)).elementAt(0); |
---|
875 | return root.makeTree(BestFirstElements, root, train, |
---|
876 | nextNode.m_SortedIndices, nextNode.m_Weights, nextNode.m_Dists, |
---|
877 | nextNode.m_ClassProbs, nextNode.m_TotalWeight, |
---|
878 | nextNode.m_Props, minNumObj, useHeuristic, useGini); |
---|
879 | } |
---|
880 | |
---|
881 | // split the node |
---|
882 | else { |
---|
883 | nodeToSplit.m_isLeaf = false; |
---|
884 | nodeToSplit.m_Attribute = att; |
---|
885 | |
---|
886 | nodeToSplit.makeSuccessors(BestFirstElements,train,subsetIndices, |
---|
887 | subsetWeights,dists, nodeToSplit.m_Attribute,useHeuristic,useGini); |
---|
888 | |
---|
889 | for (int i=0; i<2; i++){ |
---|
890 | nodeToSplit.m_Successors[i].makeLeaf(train); |
---|
891 | } |
---|
892 | |
---|
893 | return true; |
---|
894 | } |
---|
895 | } |
---|
896 | } |
---|
897 | |
---|
898 | /** |
---|
899 | * This method is to find the number of expansions based on internal |
---|
900 | * cross-validation for just post-pruning. It expands the first BestFirst |
---|
901 | * node in the BestFirstElements until no node can be split. When building |
---|
902 | * the tree, stroe error for each temporary tree, namely for each expansion. |
---|
903 | * |
---|
904 | * @param BestFirstElements list to store BFTree nodes |
---|
905 | * @param root root node of tree in each fold |
---|
906 | * @param train training data in each fold |
---|
907 | * @param test test data in each fold |
---|
908 | * @param modelError list to store error for each expansion in |
---|
909 | * each fold |
---|
910 | * @param sortedIndices sorted indices of the instances |
---|
911 | * @param weights weights of the instances |
---|
912 | * @param dists class distributions for each attribute |
---|
913 | * @param classProbs class probabilities of this node |
---|
914 | * @param totalWeight total weight of this node (note if the node |
---|
915 | * can not split, this value is not calculated.) |
---|
916 | * @param branchProps proportions of two subbranches |
---|
917 | * @param minNumObj minimal number of instances at leaf nodes |
---|
918 | * @param useHeuristic if use heuristic search for nominal attributes |
---|
919 | * in multi-class problem |
---|
920 | * @param useGini if use Gini index as splitting criterion |
---|
921 | * @param useErrorRate if use error rate in internal cross-validation |
---|
922 | * @throws Exception if something goes wrong |
---|
923 | */ |
---|
924 | protected void makeTree(FastVector BestFirstElements, BFTree root, |
---|
925 | Instances train, Instances test, FastVector modelError, int[][] sortedIndices, |
---|
926 | double[][] weights, double[][][] dists, double[] classProbs, double totalWeight, |
---|
927 | double[] branchProps, int minNumObj, boolean useHeuristic, boolean useGini, boolean useErrorRate) |
---|
928 | throws Exception { |
---|
929 | |
---|
930 | if (BestFirstElements.size()==0) return; |
---|
931 | |
---|
932 | /////////////////////////////////////////////////////////////////////// |
---|
933 | // All information about the node to split (first BestFirst object in |
---|
934 | // BestFirstElements) |
---|
935 | FastVector firstElement = (FastVector)BestFirstElements.elementAt(0); |
---|
936 | |
---|
937 | // node to split |
---|
938 | //BFTree nodeToSplit = (BFTree)firstElement.elementAt(0); |
---|
939 | |
---|
940 | // split attribute |
---|
941 | Attribute att = (Attribute)firstElement.elementAt(1); |
---|
942 | |
---|
943 | // info of split value or split string |
---|
944 | double splitValue = Double.NaN; |
---|
945 | String splitStr = null; |
---|
946 | if (att.isNumeric()) |
---|
947 | splitValue = ((Double)firstElement.elementAt(2)).doubleValue(); |
---|
948 | else { |
---|
949 | splitStr=((String)firstElement.elementAt(2)).toString(); |
---|
950 | } |
---|
951 | |
---|
952 | // the best gini gain or information of this node |
---|
953 | double gain = ((Double)firstElement.elementAt(3)).doubleValue(); |
---|
954 | /////////////////////////////////////////////////////////////////////// |
---|
955 | |
---|
956 | if (totalWeight < 2*minNumObj || branchProps[0]==0 |
---|
957 | || branchProps[1]==0) { |
---|
958 | // remove the first element |
---|
959 | BestFirstElements.removeElementAt(0); |
---|
960 | makeLeaf(train); |
---|
961 | if (BestFirstElements.size() == 0) { |
---|
962 | return; |
---|
963 | } |
---|
964 | |
---|
965 | BFTree nextSplitNode = (BFTree) |
---|
966 | ((FastVector)BestFirstElements.elementAt(0)).elementAt(0); |
---|
967 | nextSplitNode.makeTree(BestFirstElements, root, train, test, modelError, |
---|
968 | nextSplitNode.m_SortedIndices, nextSplitNode.m_Weights, |
---|
969 | nextSplitNode.m_Dists, nextSplitNode.m_ClassProbs, |
---|
970 | nextSplitNode.m_TotalWeight, nextSplitNode.m_Props, minNumObj, |
---|
971 | useHeuristic, useGini, useErrorRate); |
---|
972 | return; |
---|
973 | |
---|
974 | } |
---|
975 | |
---|
976 | // If gini gain or information gain is 0, make all nodes in the BestFirstElements leaf nodes |
---|
977 | // because these node sorted descendingly according to gini gain or information gain. |
---|
978 | // (namely, gini gain or information gain of all nodes in BestFirstEelements is 0). |
---|
979 | if (gain==0) { |
---|
980 | for (int i=0; i<BestFirstElements.size(); i++) { |
---|
981 | FastVector element = (FastVector)BestFirstElements.elementAt(i); |
---|
982 | BFTree node = (BFTree)element.elementAt(0); |
---|
983 | node.makeLeaf(train); |
---|
984 | } |
---|
985 | BestFirstElements.removeAllElements(); |
---|
986 | } |
---|
987 | |
---|
988 | // gini gain or information gain is not 0 |
---|
989 | else { |
---|
990 | // remove the first element |
---|
991 | BestFirstElements.removeElementAt(0); |
---|
992 | m_Attribute = att; |
---|
993 | if (att.isNumeric()) m_SplitValue = splitValue; |
---|
994 | else m_SplitString = splitStr; |
---|
995 | |
---|
996 | int[][][] subsetIndices = new int[2][train.numAttributes()][0]; |
---|
997 | double[][][] subsetWeights = new double[2][train.numAttributes()][0]; |
---|
998 | |
---|
999 | splitData(subsetIndices, subsetWeights, m_Attribute, |
---|
1000 | m_SplitValue, m_SplitString, |
---|
1001 | sortedIndices, weights, train); |
---|
1002 | |
---|
1003 | // if split will generate node(s) which has total weights less than m_minNumObj, |
---|
1004 | // do not split |
---|
1005 | int attIndex = att.index(); |
---|
1006 | if (subsetIndices[0][attIndex].length<minNumObj || |
---|
1007 | subsetIndices[1][attIndex].length<minNumObj) { |
---|
1008 | makeLeaf(train); |
---|
1009 | } |
---|
1010 | |
---|
1011 | // split the node and cauculate error rate of this temporary tree |
---|
1012 | else { |
---|
1013 | m_isLeaf = false; |
---|
1014 | m_Attribute = att; |
---|
1015 | |
---|
1016 | makeSuccessors(BestFirstElements,train,subsetIndices, |
---|
1017 | subsetWeights,dists, m_Attribute, useHeuristic, useGini); |
---|
1018 | for (int i=0; i<2; i++){ |
---|
1019 | m_Successors[i].makeLeaf(train); |
---|
1020 | } |
---|
1021 | |
---|
1022 | Evaluation eval = new Evaluation(test); |
---|
1023 | eval.evaluateModel(root, test); |
---|
1024 | double error; |
---|
1025 | if (useErrorRate) error = eval.errorRate(); |
---|
1026 | else error = eval.rootMeanSquaredError(); |
---|
1027 | modelError.addElement(new Double(error)); |
---|
1028 | } |
---|
1029 | |
---|
1030 | if (BestFirstElements.size()!=0) { |
---|
1031 | FastVector nextSplitElement = (FastVector)BestFirstElements.elementAt(0); |
---|
1032 | BFTree nextSplitNode = (BFTree)nextSplitElement.elementAt(0); |
---|
1033 | nextSplitNode.makeTree(BestFirstElements, root, train, test, modelError, |
---|
1034 | nextSplitNode.m_SortedIndices, nextSplitNode.m_Weights, |
---|
1035 | nextSplitNode.m_Dists, nextSplitNode.m_ClassProbs, |
---|
1036 | nextSplitNode.m_TotalWeight, nextSplitNode.m_Props, minNumObj, |
---|
1037 | useHeuristic, useGini,useErrorRate); |
---|
1038 | } |
---|
1039 | } |
---|
1040 | } |
---|
1041 | |
---|
1042 | |
---|
1043 | /** |
---|
1044 | * Generate successor nodes for a node and put them into BestFirstElements |
---|
1045 | * according to gini gain or information gain in a descending order. |
---|
1046 | * |
---|
1047 | * @param BestFirstElements list to store BestFirst nodes |
---|
1048 | * @param data training instance |
---|
1049 | * @param subsetSortedIndices sorted indices of instances of successor nodes |
---|
1050 | * @param subsetWeights weights of instances of successor nodes |
---|
1051 | * @param dists class distributions of successor nodes |
---|
1052 | * @param att attribute used to split the node |
---|
1053 | * @param useHeuristic if use heuristic search for nominal attributes in multi-class problem |
---|
1054 | * @param useGini if use Gini index as splitting criterion |
---|
1055 | * @throws Exception if something goes wrong |
---|
1056 | */ |
---|
1057 | protected void makeSuccessors(FastVector BestFirstElements,Instances data, |
---|
1058 | int[][][] subsetSortedIndices, double[][][] subsetWeights, |
---|
1059 | double[][][] dists, |
---|
1060 | Attribute att, boolean useHeuristic, boolean useGini) throws Exception { |
---|
1061 | |
---|
1062 | m_Successors = new BFTree[2]; |
---|
1063 | |
---|
1064 | for (int i=0; i<2; i++) { |
---|
1065 | m_Successors[i] = new BFTree(); |
---|
1066 | m_Successors[i].m_isLeaf = true; |
---|
1067 | |
---|
1068 | // class probability and distribution for this successor node |
---|
1069 | m_Successors[i].m_ClassProbs = new double[data.numClasses()]; |
---|
1070 | m_Successors[i].m_Distribution = new double[data.numClasses()]; |
---|
1071 | System.arraycopy(dists[att.index()][i], 0, m_Successors[i].m_ClassProbs, |
---|
1072 | 0,m_Successors[i].m_ClassProbs.length); |
---|
1073 | System.arraycopy(dists[att.index()][i], 0, m_Successors[i].m_Distribution, |
---|
1074 | 0,m_Successors[i].m_Distribution.length); |
---|
1075 | if (Utils.sum(m_Successors[i].m_ClassProbs)!=0) |
---|
1076 | Utils.normalize(m_Successors[i].m_ClassProbs); |
---|
1077 | |
---|
1078 | // split information for this successor node |
---|
1079 | double[][] props = new double[data.numAttributes()][2]; |
---|
1080 | double[][][] subDists = new double[data.numAttributes()][2][data.numClasses()]; |
---|
1081 | double[][] totalSubsetWeights = new double[data.numAttributes()][2]; |
---|
1082 | FastVector splitInfo = m_Successors[i].computeSplitInfo(m_Successors[i], data, |
---|
1083 | subsetSortedIndices[i], subsetWeights[i], subDists, props, |
---|
1084 | totalSubsetWeights, useHeuristic, useGini); |
---|
1085 | |
---|
1086 | // branch proportion for this successor node |
---|
1087 | int splitIndex = ((Attribute)splitInfo.elementAt(1)).index(); |
---|
1088 | m_Successors[i].m_Props = new double[2]; |
---|
1089 | System.arraycopy(props[splitIndex], 0, m_Successors[i].m_Props, 0, |
---|
1090 | m_Successors[i].m_Props.length); |
---|
1091 | |
---|
1092 | // sorted indices and weights of each attribute for this successor node |
---|
1093 | m_Successors[i].m_SortedIndices = new int[data.numAttributes()][0]; |
---|
1094 | m_Successors[i].m_Weights = new double[data.numAttributes()][0]; |
---|
1095 | for (int j=0; j<m_Successors[i].m_SortedIndices.length; j++) { |
---|
1096 | m_Successors[i].m_SortedIndices[j] = subsetSortedIndices[i][j]; |
---|
1097 | m_Successors[i].m_Weights[j] = subsetWeights[i][j]; |
---|
1098 | } |
---|
1099 | |
---|
1100 | // distribution of each attribute for this successor node |
---|
1101 | m_Successors[i].m_Dists = new double[data.numAttributes()][2][data.numClasses()]; |
---|
1102 | for (int j=0; j<subDists.length; j++) { |
---|
1103 | m_Successors[i].m_Dists[j] = subDists[j]; |
---|
1104 | } |
---|
1105 | |
---|
1106 | // total weights for this successor node. |
---|
1107 | m_Successors[i].m_TotalWeight = Utils.sum(totalSubsetWeights[splitIndex]); |
---|
1108 | |
---|
1109 | // insert this successor node into BestFirstElements according to gini gain or information gain |
---|
1110 | // descendingly |
---|
1111 | if (BestFirstElements.size()==0) { |
---|
1112 | BestFirstElements.addElement(splitInfo); |
---|
1113 | } else { |
---|
1114 | double gGain = ((Double)(splitInfo.elementAt(3))).doubleValue(); |
---|
1115 | int vectorSize = BestFirstElements.size(); |
---|
1116 | FastVector lastNode = (FastVector)BestFirstElements.elementAt(vectorSize-1); |
---|
1117 | |
---|
1118 | // If gini gain is less than that of last node in FastVector |
---|
1119 | if (gGain<((Double)(lastNode.elementAt(3))).doubleValue()) { |
---|
1120 | BestFirstElements.insertElementAt(splitInfo, vectorSize); |
---|
1121 | } else { |
---|
1122 | for (int j=0; j<vectorSize; j++) { |
---|
1123 | FastVector node = (FastVector)BestFirstElements.elementAt(j); |
---|
1124 | double nodeGain = ((Double)(node.elementAt(3))).doubleValue(); |
---|
1125 | if (gGain>=nodeGain) { |
---|
1126 | BestFirstElements.insertElementAt(splitInfo, j); |
---|
1127 | break; |
---|
1128 | } |
---|
1129 | } |
---|
1130 | } |
---|
1131 | } |
---|
1132 | } |
---|
1133 | } |
---|
1134 | |
---|
1135 | /** |
---|
1136 | * Compute sorted indices, weights and class probabilities for a given |
---|
1137 | * dataset. Return total weights of the data at the node. |
---|
1138 | * |
---|
1139 | * @param data training data |
---|
1140 | * @param sortedIndices sorted indices of instances at the node |
---|
1141 | * @param weights weights of instances at the node |
---|
1142 | * @param classProbs class probabilities at the node |
---|
1143 | * @return total weights of instances at the node |
---|
1144 | * @throws Exception if something goes wrong |
---|
1145 | */ |
---|
1146 | protected double computeSortedInfo(Instances data, int[][] sortedIndices, double[][] weights, |
---|
1147 | double[] classProbs) throws Exception { |
---|
1148 | |
---|
1149 | // Create array of sorted indices and weights |
---|
1150 | double[] vals = new double[data.numInstances()]; |
---|
1151 | for (int j = 0; j < data.numAttributes(); j++) { |
---|
1152 | if (j==data.classIndex()) continue; |
---|
1153 | weights[j] = new double[data.numInstances()]; |
---|
1154 | |
---|
1155 | if (data.attribute(j).isNominal()) { |
---|
1156 | |
---|
1157 | // Handling nominal attributes. Putting indices of |
---|
1158 | // instances with missing values at the end. |
---|
1159 | sortedIndices[j] = new int[data.numInstances()]; |
---|
1160 | int count = 0; |
---|
1161 | for (int i = 0; i < data.numInstances(); i++) { |
---|
1162 | Instance inst = data.instance(i); |
---|
1163 | if (!inst.isMissing(j)) { |
---|
1164 | sortedIndices[j][count] = i; |
---|
1165 | weights[j][count] = inst.weight(); |
---|
1166 | count++; |
---|
1167 | } |
---|
1168 | } |
---|
1169 | for (int i = 0; i < data.numInstances(); i++) { |
---|
1170 | Instance inst = data.instance(i); |
---|
1171 | if (inst.isMissing(j)) { |
---|
1172 | sortedIndices[j][count] = i; |
---|
1173 | weights[j][count] = inst.weight(); |
---|
1174 | count++; |
---|
1175 | } |
---|
1176 | } |
---|
1177 | } else { |
---|
1178 | |
---|
1179 | // Sorted indices are computed for numeric attributes |
---|
1180 | // missing values instances are put to end (through Utils.sort() method) |
---|
1181 | for (int i = 0; i < data.numInstances(); i++) { |
---|
1182 | Instance inst = data.instance(i); |
---|
1183 | vals[i] = inst.value(j); |
---|
1184 | } |
---|
1185 | sortedIndices[j] = Utils.sort(vals); |
---|
1186 | for (int i = 0; i < data.numInstances(); i++) { |
---|
1187 | weights[j][i] = data.instance(sortedIndices[j][i]).weight(); |
---|
1188 | } |
---|
1189 | } |
---|
1190 | } |
---|
1191 | |
---|
1192 | // Compute initial class counts and total weight |
---|
1193 | double totalWeight = 0; |
---|
1194 | for (int i = 0; i < data.numInstances(); i++) { |
---|
1195 | Instance inst = data.instance(i); |
---|
1196 | classProbs[(int)inst.classValue()] += inst.weight(); |
---|
1197 | totalWeight += inst.weight(); |
---|
1198 | } |
---|
1199 | |
---|
1200 | return totalWeight; |
---|
1201 | } |
---|
1202 | |
---|
1203 | /** |
---|
1204 | * Compute the best splitting attribute, split point or subset and the best |
---|
1205 | * gini gain or iformation gain for a given dataset. |
---|
1206 | * |
---|
1207 | * @param node node to be split |
---|
1208 | * @param data training data |
---|
1209 | * @param sortedIndices sorted indices of the instances |
---|
1210 | * @param weights weights of the instances |
---|
1211 | * @param dists class distributions for each attribute |
---|
1212 | * @param props proportions of two branches |
---|
1213 | * @param totalSubsetWeights total weight of two subsets |
---|
1214 | * @param useHeuristic if use heuristic search for nominal attributes |
---|
1215 | * in multi-class problem |
---|
1216 | * @param useGini if use Gini index as splitting criterion |
---|
1217 | * @return split information about the node |
---|
1218 | * @throws Exception if something is wrong |
---|
1219 | */ |
---|
1220 | protected FastVector computeSplitInfo(BFTree node, Instances data, int[][] sortedIndices, |
---|
1221 | double[][] weights, double[][][] dists, double[][] props, |
---|
1222 | double[][] totalSubsetWeights, boolean useHeuristic, boolean useGini) throws Exception { |
---|
1223 | |
---|
1224 | double[] splits = new double[data.numAttributes()]; |
---|
1225 | String[] splitString = new String[data.numAttributes()]; |
---|
1226 | double[] gains = new double[data.numAttributes()]; |
---|
1227 | |
---|
1228 | for (int i = 0; i < data.numAttributes(); i++) { |
---|
1229 | if (i==data.classIndex()) continue; |
---|
1230 | Attribute att = data.attribute(i); |
---|
1231 | if (att.isNumeric()) { |
---|
1232 | // numeric attribute |
---|
1233 | splits[i] = numericDistribution(props, dists, att, sortedIndices[i], |
---|
1234 | weights[i], totalSubsetWeights, gains, data, useGini); |
---|
1235 | } else { |
---|
1236 | // nominal attribute |
---|
1237 | splitString[i] = nominalDistribution(props, dists, att, sortedIndices[i], |
---|
1238 | weights[i], totalSubsetWeights, gains, data, useHeuristic, useGini); |
---|
1239 | } |
---|
1240 | } |
---|
1241 | |
---|
1242 | int index = Utils.maxIndex(gains); |
---|
1243 | double mBestGain = gains[index]; |
---|
1244 | |
---|
1245 | Attribute att = data.attribute(index); |
---|
1246 | double mValue =Double.NaN; |
---|
1247 | String mString = null; |
---|
1248 | if (att.isNumeric()) mValue= splits[index]; |
---|
1249 | else { |
---|
1250 | mString = splitString[index]; |
---|
1251 | if (mString==null) mString = ""; |
---|
1252 | } |
---|
1253 | |
---|
1254 | // split information |
---|
1255 | FastVector splitInfo = new FastVector(); |
---|
1256 | splitInfo.addElement(node); |
---|
1257 | splitInfo.addElement(att); |
---|
1258 | if (att.isNumeric()) splitInfo.addElement(new Double(mValue)); |
---|
1259 | else splitInfo.addElement(mString); |
---|
1260 | splitInfo.addElement(new Double(mBestGain)); |
---|
1261 | |
---|
1262 | return splitInfo; |
---|
1263 | } |
---|
1264 | |
---|
1265 | /** |
---|
1266 | * Compute distributions, proportions and total weights of two successor nodes for |
---|
1267 | * a given numeric attribute. |
---|
1268 | * |
---|
1269 | * @param props proportions of each two branches for each attribute |
---|
1270 | * @param dists class distributions of two branches for each attribute |
---|
1271 | * @param att numeric att split on |
---|
1272 | * @param sortedIndices sorted indices of instances for the attirubte |
---|
1273 | * @param weights weights of instances for the attirbute |
---|
1274 | * @param subsetWeights total weight of two branches split based on the attribute |
---|
1275 | * @param gains Gini gains or information gains for each attribute |
---|
1276 | * @param data training instances |
---|
1277 | * @param useGini if use Gini index as splitting criterion |
---|
1278 | * @return Gini gain or information gain for the given attribute |
---|
1279 | * @throws Exception if something goes wrong |
---|
1280 | */ |
---|
1281 | protected double numericDistribution(double[][] props, double[][][] dists, |
---|
1282 | Attribute att, int[] sortedIndices, double[] weights, double[][] subsetWeights, |
---|
1283 | double[] gains, Instances data, boolean useGini) |
---|
1284 | throws Exception { |
---|
1285 | |
---|
1286 | double splitPoint = Double.NaN; |
---|
1287 | double[][] dist = null; |
---|
1288 | int numClasses = data.numClasses(); |
---|
1289 | int i; // differ instances with or without missing values |
---|
1290 | |
---|
1291 | double[][] currDist = new double[2][numClasses]; |
---|
1292 | dist = new double[2][numClasses]; |
---|
1293 | |
---|
1294 | // Move all instances without missing values into second subset |
---|
1295 | double[] parentDist = new double[numClasses]; |
---|
1296 | int missingStart = 0; |
---|
1297 | for (int j = 0; j < sortedIndices.length; j++) { |
---|
1298 | Instance inst = data.instance(sortedIndices[j]); |
---|
1299 | if (!inst.isMissing(att)) { |
---|
1300 | missingStart ++; |
---|
1301 | currDist[1][(int)inst.classValue()] += weights[j]; |
---|
1302 | } |
---|
1303 | parentDist[(int)inst.classValue()] += weights[j]; |
---|
1304 | } |
---|
1305 | System.arraycopy(currDist[1], 0, dist[1], 0, dist[1].length); |
---|
1306 | |
---|
1307 | // Try all possible split points |
---|
1308 | double currSplit = data.instance(sortedIndices[0]).value(att); |
---|
1309 | double currGain; |
---|
1310 | double bestGain = -Double.MAX_VALUE; |
---|
1311 | |
---|
1312 | for (i = 0; i < sortedIndices.length; i++) { |
---|
1313 | Instance inst = data.instance(sortedIndices[i]); |
---|
1314 | if (inst.isMissing(att)) { |
---|
1315 | break; |
---|
1316 | } |
---|
1317 | if (inst.value(att) > currSplit) { |
---|
1318 | |
---|
1319 | double[][] tempDist = new double[2][numClasses]; |
---|
1320 | for (int k=0; k<2; k++) { |
---|
1321 | //tempDist[k] = currDist[k]; |
---|
1322 | System.arraycopy(currDist[k], 0, tempDist[k], 0, tempDist[k].length); |
---|
1323 | } |
---|
1324 | |
---|
1325 | double[] tempProps = new double[2]; |
---|
1326 | for (int k=0; k<2; k++) { |
---|
1327 | tempProps[k] = Utils.sum(tempDist[k]); |
---|
1328 | } |
---|
1329 | |
---|
1330 | if (Utils.sum(tempProps) !=0) Utils.normalize(tempProps); |
---|
1331 | |
---|
1332 | // split missing values |
---|
1333 | int index = missingStart; |
---|
1334 | while (index < sortedIndices.length) { |
---|
1335 | Instance insta = data.instance(sortedIndices[index]); |
---|
1336 | for (int j = 0; j < 2; j++) { |
---|
1337 | tempDist[j][(int)insta.classValue()] += tempProps[j] * weights[index]; |
---|
1338 | } |
---|
1339 | index++; |
---|
1340 | } |
---|
1341 | |
---|
1342 | if (useGini) currGain = computeGiniGain(parentDist,tempDist); |
---|
1343 | else currGain = computeInfoGain(parentDist,tempDist); |
---|
1344 | |
---|
1345 | if (currGain > bestGain) { |
---|
1346 | bestGain = currGain; |
---|
1347 | // clean split point |
---|
1348 | splitPoint = Math.rint((inst.value(att) + currSplit)/2.0*100000)/100000.0; |
---|
1349 | for (int j = 0; j < currDist.length; j++) { |
---|
1350 | System.arraycopy(tempDist[j], 0, dist[j], 0, |
---|
1351 | dist[j].length); |
---|
1352 | } |
---|
1353 | } |
---|
1354 | } |
---|
1355 | currSplit = inst.value(att); |
---|
1356 | currDist[0][(int)inst.classValue()] += weights[i]; |
---|
1357 | currDist[1][(int)inst.classValue()] -= weights[i]; |
---|
1358 | } |
---|
1359 | |
---|
1360 | // Compute weights |
---|
1361 | int attIndex = att.index(); |
---|
1362 | props[attIndex] = new double[2]; |
---|
1363 | for (int k = 0; k < 2; k++) { |
---|
1364 | props[attIndex][k] = Utils.sum(dist[k]); |
---|
1365 | } |
---|
1366 | if (Utils.sum(props[attIndex]) != 0) Utils.normalize(props[attIndex]); |
---|
1367 | |
---|
1368 | // Compute subset weights |
---|
1369 | subsetWeights[attIndex] = new double[2]; |
---|
1370 | for (int j = 0; j < 2; j++) { |
---|
1371 | subsetWeights[attIndex][j] += Utils.sum(dist[j]); |
---|
1372 | } |
---|
1373 | |
---|
1374 | // clean gain |
---|
1375 | gains[attIndex] = Math.rint(bestGain*10000000)/10000000.0; |
---|
1376 | dists[attIndex] = dist; |
---|
1377 | return splitPoint; |
---|
1378 | } |
---|
1379 | |
---|
1380 | /** |
---|
1381 | * Compute distributions, proportions and total weights of two successor |
---|
1382 | * nodes for a given nominal attribute. |
---|
1383 | * |
---|
1384 | * @param props proportions of each two branches for each attribute |
---|
1385 | * @param dists class distributions of two branches for each attribute |
---|
1386 | * @param att numeric att split on |
---|
1387 | * @param sortedIndices sorted indices of instances for the attirubte |
---|
1388 | * @param weights weights of instances for the attirbute |
---|
1389 | * @param subsetWeights total weight of two branches split based on the attribute |
---|
1390 | * @param gains Gini gains for each attribute |
---|
1391 | * @param data training instances |
---|
1392 | * @param useHeuristic if use heuristic search |
---|
1393 | * @param useGini if use Gini index as splitting criterion |
---|
1394 | * @return Gini gain for the given attribute |
---|
1395 | * @throws Exception if something goes wrong |
---|
1396 | */ |
---|
1397 | protected String nominalDistribution(double[][] props, double[][][] dists, |
---|
1398 | Attribute att, int[] sortedIndices, double[] weights, double[][] subsetWeights, |
---|
1399 | double[] gains, Instances data, boolean useHeuristic, boolean useGini) |
---|
1400 | throws Exception { |
---|
1401 | |
---|
1402 | String[] values = new String[att.numValues()]; |
---|
1403 | int numCat = values.length; // number of values of the attribute |
---|
1404 | int numClasses = data.numClasses(); |
---|
1405 | |
---|
1406 | String bestSplitString = ""; |
---|
1407 | double bestGain = -Double.MAX_VALUE; |
---|
1408 | |
---|
1409 | // class frequency for each value |
---|
1410 | int[] classFreq = new int[numCat]; |
---|
1411 | for (int j=0; j<numCat; j++) classFreq[j] = 0; |
---|
1412 | |
---|
1413 | double[] parentDist = new double[numClasses]; |
---|
1414 | double[][] currDist = new double[2][numClasses]; |
---|
1415 | double[][] dist = new double[2][numClasses]; |
---|
1416 | int missingStart = 0; |
---|
1417 | |
---|
1418 | for (int i = 0; i < sortedIndices.length; i++) { |
---|
1419 | Instance inst = data.instance(sortedIndices[i]); |
---|
1420 | if (!inst.isMissing(att)) { |
---|
1421 | missingStart++; |
---|
1422 | classFreq[(int)inst.value(att)] ++; |
---|
1423 | } |
---|
1424 | parentDist[(int)inst.classValue()] += weights[i]; |
---|
1425 | } |
---|
1426 | |
---|
1427 | // count the number of values that class frequency is not 0 |
---|
1428 | int nonEmpty = 0; |
---|
1429 | for (int j=0; j<numCat; j++) { |
---|
1430 | if (classFreq[j]!=0) nonEmpty ++; |
---|
1431 | } |
---|
1432 | |
---|
1433 | // attribute values which class frequency is not 0 |
---|
1434 | String[] nonEmptyValues = new String[nonEmpty]; |
---|
1435 | int nonEmptyIndex = 0; |
---|
1436 | for (int j=0; j<numCat; j++) { |
---|
1437 | if (classFreq[j]!=0) { |
---|
1438 | nonEmptyValues[nonEmptyIndex] = att.value(j); |
---|
1439 | nonEmptyIndex ++; |
---|
1440 | } |
---|
1441 | } |
---|
1442 | |
---|
1443 | // attribute values which class frequency is 0 |
---|
1444 | int empty = numCat - nonEmpty; |
---|
1445 | String[] emptyValues = new String[empty]; |
---|
1446 | int emptyIndex = 0; |
---|
1447 | for (int j=0; j<numCat; j++) { |
---|
1448 | if (classFreq[j]==0) { |
---|
1449 | emptyValues[emptyIndex] = att.value(j); |
---|
1450 | emptyIndex ++; |
---|
1451 | } |
---|
1452 | } |
---|
1453 | |
---|
1454 | if (nonEmpty<=1) { |
---|
1455 | gains[att.index()] = 0; |
---|
1456 | return ""; |
---|
1457 | } |
---|
1458 | |
---|
1459 | // for tow-class probloms |
---|
1460 | if (data.numClasses()==2) { |
---|
1461 | |
---|
1462 | //// Firstly, for attribute values which class frequency is not zero |
---|
1463 | |
---|
1464 | // probability of class 0 for each attribute value |
---|
1465 | double[] pClass0 = new double[nonEmpty]; |
---|
1466 | // class distribution for each attribute value |
---|
1467 | double[][] valDist = new double[nonEmpty][2]; |
---|
1468 | |
---|
1469 | for (int j=0; j<nonEmpty; j++) { |
---|
1470 | for (int k=0; k<2; k++) { |
---|
1471 | valDist[j][k] = 0; |
---|
1472 | } |
---|
1473 | } |
---|
1474 | |
---|
1475 | for (int i = 0; i < sortedIndices.length; i++) { |
---|
1476 | Instance inst = data.instance(sortedIndices[i]); |
---|
1477 | if (inst.isMissing(att)) { |
---|
1478 | break; |
---|
1479 | } |
---|
1480 | |
---|
1481 | for (int j=0; j<nonEmpty; j++) { |
---|
1482 | if (att.value((int)inst.value(att)).compareTo(nonEmptyValues[j])==0) { |
---|
1483 | valDist[j][(int)inst.classValue()] += inst.weight(); |
---|
1484 | break; |
---|
1485 | } |
---|
1486 | } |
---|
1487 | } |
---|
1488 | |
---|
1489 | for (int j=0; j<nonEmpty; j++) { |
---|
1490 | double distSum = Utils.sum(valDist[j]); |
---|
1491 | if (distSum==0) pClass0[j]=0; |
---|
1492 | else pClass0[j] = valDist[j][0]/distSum; |
---|
1493 | } |
---|
1494 | |
---|
1495 | // sort category according to the probability of class 0.0 |
---|
1496 | String[] sortedValues = new String[nonEmpty]; |
---|
1497 | for (int j=0; j<nonEmpty; j++) { |
---|
1498 | sortedValues[j] = nonEmptyValues[Utils.minIndex(pClass0)]; |
---|
1499 | pClass0[Utils.minIndex(pClass0)] = Double.MAX_VALUE; |
---|
1500 | } |
---|
1501 | |
---|
1502 | // Find a subset of attribute values that maximize impurity decrease |
---|
1503 | |
---|
1504 | // for the attribute values that class frequency is not 0 |
---|
1505 | String tempStr = ""; |
---|
1506 | |
---|
1507 | for (int j=0; j<nonEmpty-1; j++) { |
---|
1508 | currDist = new double[2][numClasses]; |
---|
1509 | if (tempStr=="") tempStr="(" + sortedValues[j] + ")"; |
---|
1510 | else tempStr += "|"+ "(" + sortedValues[j] + ")"; |
---|
1511 | //System.out.println(sortedValues[j]); |
---|
1512 | for (int i=0; i<sortedIndices.length;i++) { |
---|
1513 | Instance inst = data.instance(sortedIndices[i]); |
---|
1514 | if (inst.isMissing(att)) { |
---|
1515 | break; |
---|
1516 | } |
---|
1517 | |
---|
1518 | if (tempStr.indexOf |
---|
1519 | ("(" + att.value((int)inst.value(att)) + ")")!=-1) { |
---|
1520 | currDist[0][(int)inst.classValue()] += weights[i]; |
---|
1521 | } else currDist[1][(int)inst.classValue()] += weights[i]; |
---|
1522 | } |
---|
1523 | |
---|
1524 | double[][] tempDist = new double[2][numClasses]; |
---|
1525 | for (int kk=0; kk<2; kk++) { |
---|
1526 | tempDist[kk] = currDist[kk]; |
---|
1527 | } |
---|
1528 | |
---|
1529 | double[] tempProps = new double[2]; |
---|
1530 | for (int kk=0; kk<2; kk++) { |
---|
1531 | tempProps[kk] = Utils.sum(tempDist[kk]); |
---|
1532 | } |
---|
1533 | |
---|
1534 | if (Utils.sum(tempProps)!=0) Utils.normalize(tempProps); |
---|
1535 | |
---|
1536 | // split missing values |
---|
1537 | int mstart = missingStart; |
---|
1538 | while (mstart < sortedIndices.length) { |
---|
1539 | Instance insta = data.instance(sortedIndices[mstart]); |
---|
1540 | for (int jj = 0; jj < 2; jj++) { |
---|
1541 | tempDist[jj][(int)insta.classValue()] += tempProps[jj] * weights[mstart]; |
---|
1542 | } |
---|
1543 | mstart++; |
---|
1544 | } |
---|
1545 | |
---|
1546 | double currGain; |
---|
1547 | if (useGini) currGain = computeGiniGain(parentDist,tempDist); |
---|
1548 | else currGain = computeInfoGain(parentDist,tempDist); |
---|
1549 | |
---|
1550 | if (currGain>bestGain) { |
---|
1551 | bestGain = currGain; |
---|
1552 | bestSplitString = tempStr; |
---|
1553 | for (int jj = 0; jj < 2; jj++) { |
---|
1554 | System.arraycopy(tempDist[jj], 0, dist[jj], 0, |
---|
1555 | dist[jj].length); |
---|
1556 | } |
---|
1557 | } |
---|
1558 | } |
---|
1559 | } |
---|
1560 | |
---|
1561 | // multi-class problems (exhaustive search) |
---|
1562 | else if (!useHeuristic || nonEmpty<=4) { |
---|
1563 | //else if (!useHeuristic || nonEmpty==2) { |
---|
1564 | |
---|
1565 | // Firstly, for attribute values which class frequency is not zero |
---|
1566 | for (int i=0; i<(int)Math.pow(2,nonEmpty-1); i++) { |
---|
1567 | String tempStr=""; |
---|
1568 | currDist = new double[2][numClasses]; |
---|
1569 | int mod; |
---|
1570 | int bit10 = i; |
---|
1571 | for (int j=nonEmpty-1; j>=0; j--) { |
---|
1572 | mod = bit10%2; // convert from 10bit to 2bit |
---|
1573 | if (mod==1) { |
---|
1574 | if (tempStr=="") tempStr = "("+nonEmptyValues[j]+")"; |
---|
1575 | else tempStr += "|" + "("+nonEmptyValues[j]+")"; |
---|
1576 | } |
---|
1577 | bit10 = bit10/2; |
---|
1578 | } |
---|
1579 | for (int j=0; j<sortedIndices.length;j++) { |
---|
1580 | Instance inst = data.instance(sortedIndices[j]); |
---|
1581 | if (inst.isMissing(att)) { |
---|
1582 | break; |
---|
1583 | } |
---|
1584 | |
---|
1585 | if (tempStr.indexOf("("+att.value((int)inst.value(att))+")")!=-1) { |
---|
1586 | currDist[0][(int)inst.classValue()] += weights[j]; |
---|
1587 | } else currDist[1][(int)inst.classValue()] += weights[j]; |
---|
1588 | } |
---|
1589 | |
---|
1590 | double[][] tempDist = new double[2][numClasses]; |
---|
1591 | for (int k=0; k<2; k++) { |
---|
1592 | tempDist[k] = currDist[k]; |
---|
1593 | } |
---|
1594 | |
---|
1595 | double[] tempProps = new double[2]; |
---|
1596 | for (int k=0; k<2; k++) { |
---|
1597 | tempProps[k] = Utils.sum(tempDist[k]); |
---|
1598 | } |
---|
1599 | |
---|
1600 | if (Utils.sum(tempProps)!=0) Utils.normalize(tempProps); |
---|
1601 | |
---|
1602 | // split missing values |
---|
1603 | int index = missingStart; |
---|
1604 | while (index < sortedIndices.length) { |
---|
1605 | Instance insta = data.instance(sortedIndices[index]); |
---|
1606 | for (int j = 0; j < 2; j++) { |
---|
1607 | tempDist[j][(int)insta.classValue()] += tempProps[j] * weights[index]; |
---|
1608 | } |
---|
1609 | index++; |
---|
1610 | } |
---|
1611 | |
---|
1612 | double currGain; |
---|
1613 | if (useGini) currGain = computeGiniGain(parentDist,tempDist); |
---|
1614 | else currGain = computeInfoGain(parentDist,tempDist); |
---|
1615 | |
---|
1616 | if (currGain>bestGain) { |
---|
1617 | bestGain = currGain; |
---|
1618 | bestSplitString = tempStr; |
---|
1619 | for (int j = 0; j < 2; j++) { |
---|
1620 | //dist[jj] = new double[currDist[jj].length]; |
---|
1621 | System.arraycopy(tempDist[j], 0, dist[j], 0, |
---|
1622 | dist[j].length); |
---|
1623 | } |
---|
1624 | } |
---|
1625 | } |
---|
1626 | } |
---|
1627 | |
---|
1628 | // huristic method to solve multi-classes problems |
---|
1629 | else { |
---|
1630 | // Firstly, for attribute values which class frequency is not zero |
---|
1631 | int n = nonEmpty; |
---|
1632 | int k = data.numClasses(); // number of classes of the data |
---|
1633 | double[][] P = new double[n][k]; // class probability matrix |
---|
1634 | int[] numInstancesValue = new int[n]; // number of instances for an attribute value |
---|
1635 | double[] meanClass = new double[k]; // vector of mean class probability |
---|
1636 | int numInstances = data.numInstances(); // total number of instances |
---|
1637 | |
---|
1638 | // initialize the vector of mean class probability |
---|
1639 | for (int j=0; j<meanClass.length; j++) meanClass[j]=0; |
---|
1640 | |
---|
1641 | for (int j=0; j<numInstances; j++) { |
---|
1642 | Instance inst = (Instance)data.instance(j); |
---|
1643 | int valueIndex = 0; // attribute value index in nonEmptyValues |
---|
1644 | for (int i=0; i<nonEmpty; i++) { |
---|
1645 | if (att.value((int)inst.value(att)).compareToIgnoreCase(nonEmptyValues[i])==0){ |
---|
1646 | valueIndex = i; |
---|
1647 | break; |
---|
1648 | } |
---|
1649 | } |
---|
1650 | P[valueIndex][(int)inst.classValue()]++; |
---|
1651 | numInstancesValue[valueIndex]++; |
---|
1652 | meanClass[(int)inst.classValue()]++; |
---|
1653 | } |
---|
1654 | |
---|
1655 | // calculate the class probability matrix |
---|
1656 | for (int i=0; i<P.length; i++) { |
---|
1657 | for (int j=0; j<P[0].length; j++) { |
---|
1658 | if (numInstancesValue[i]==0) P[i][j]=0; |
---|
1659 | else P[i][j]/=numInstancesValue[i]; |
---|
1660 | } |
---|
1661 | } |
---|
1662 | |
---|
1663 | //calculate the vector of mean class probability |
---|
1664 | for (int i=0; i<meanClass.length; i++) { |
---|
1665 | meanClass[i]/=numInstances; |
---|
1666 | } |
---|
1667 | |
---|
1668 | // calculate the covariance matrix |
---|
1669 | double[][] covariance = new double[k][k]; |
---|
1670 | for (int i1=0; i1<k; i1++) { |
---|
1671 | for (int i2=0; i2<k; i2++) { |
---|
1672 | double element = 0; |
---|
1673 | for (int j=0; j<n; j++) { |
---|
1674 | element += (P[j][i2]-meanClass[i2])*(P[j][i1]-meanClass[i1]) |
---|
1675 | *numInstancesValue[j]; |
---|
1676 | } |
---|
1677 | covariance[i1][i2] = element; |
---|
1678 | } |
---|
1679 | } |
---|
1680 | |
---|
1681 | Matrix matrix = new Matrix(covariance); |
---|
1682 | weka.core.matrix.EigenvalueDecomposition eigen = |
---|
1683 | new weka.core.matrix.EigenvalueDecomposition(matrix); |
---|
1684 | double[] eigenValues = eigen.getRealEigenvalues(); |
---|
1685 | |
---|
1686 | // find index of the largest eigenvalue |
---|
1687 | int index=0; |
---|
1688 | double largest = eigenValues[0]; |
---|
1689 | for (int i=1; i<eigenValues.length; i++) { |
---|
1690 | if (eigenValues[i]>largest) { |
---|
1691 | index=i; |
---|
1692 | largest = eigenValues[i]; |
---|
1693 | } |
---|
1694 | } |
---|
1695 | |
---|
1696 | // calculate the first principle component |
---|
1697 | double[] FPC = new double[k]; |
---|
1698 | Matrix eigenVector = eigen.getV(); |
---|
1699 | double[][] vectorArray = eigenVector.getArray(); |
---|
1700 | for (int i=0; i<FPC.length; i++) { |
---|
1701 | FPC[i] = vectorArray[i][index]; |
---|
1702 | } |
---|
1703 | |
---|
1704 | // calculate the first principle component scores |
---|
1705 | double[] Sa = new double[n]; |
---|
1706 | for (int i=0; i<Sa.length; i++) { |
---|
1707 | Sa[i]=0; |
---|
1708 | for (int j=0; j<k; j++) { |
---|
1709 | Sa[i] += FPC[j]*P[i][j]; |
---|
1710 | } |
---|
1711 | } |
---|
1712 | |
---|
1713 | // sort category according to Sa(s) |
---|
1714 | double[] pCopy = new double[n]; |
---|
1715 | System.arraycopy(Sa,0,pCopy,0,n); |
---|
1716 | String[] sortedValues = new String[n]; |
---|
1717 | Arrays.sort(Sa); |
---|
1718 | |
---|
1719 | for (int j=0; j<n; j++) { |
---|
1720 | sortedValues[j] = nonEmptyValues[Utils.minIndex(pCopy)]; |
---|
1721 | pCopy[Utils.minIndex(pCopy)] = Double.MAX_VALUE; |
---|
1722 | } |
---|
1723 | |
---|
1724 | // for the attribute values that class frequency is not 0 |
---|
1725 | String tempStr = ""; |
---|
1726 | |
---|
1727 | for (int j=0; j<nonEmpty-1; j++) { |
---|
1728 | currDist = new double[2][numClasses]; |
---|
1729 | if (tempStr=="") tempStr="(" + sortedValues[j] + ")"; |
---|
1730 | else tempStr += "|"+ "(" + sortedValues[j] + ")"; |
---|
1731 | for (int i=0; i<sortedIndices.length;i++) { |
---|
1732 | Instance inst = data.instance(sortedIndices[i]); |
---|
1733 | if (inst.isMissing(att)) { |
---|
1734 | break; |
---|
1735 | } |
---|
1736 | |
---|
1737 | if (tempStr.indexOf |
---|
1738 | ("(" + att.value((int)inst.value(att)) + ")")!=-1) { |
---|
1739 | currDist[0][(int)inst.classValue()] += weights[i]; |
---|
1740 | } else currDist[1][(int)inst.classValue()] += weights[i]; |
---|
1741 | } |
---|
1742 | |
---|
1743 | double[][] tempDist = new double[2][numClasses]; |
---|
1744 | for (int kk=0; kk<2; kk++) { |
---|
1745 | tempDist[kk] = currDist[kk]; |
---|
1746 | } |
---|
1747 | |
---|
1748 | double[] tempProps = new double[2]; |
---|
1749 | for (int kk=0; kk<2; kk++) { |
---|
1750 | tempProps[kk] = Utils.sum(tempDist[kk]); |
---|
1751 | } |
---|
1752 | |
---|
1753 | if (Utils.sum(tempProps)!=0) Utils.normalize(tempProps); |
---|
1754 | |
---|
1755 | // split missing values |
---|
1756 | int mstart = missingStart; |
---|
1757 | while (mstart < sortedIndices.length) { |
---|
1758 | Instance insta = data.instance(sortedIndices[mstart]); |
---|
1759 | for (int jj = 0; jj < 2; jj++) { |
---|
1760 | tempDist[jj][(int)insta.classValue()] += tempProps[jj] * weights[mstart]; |
---|
1761 | } |
---|
1762 | mstart++; |
---|
1763 | } |
---|
1764 | |
---|
1765 | double currGain; |
---|
1766 | if (useGini) currGain = computeGiniGain(parentDist,tempDist); |
---|
1767 | else currGain = computeInfoGain(parentDist,tempDist); |
---|
1768 | |
---|
1769 | if (currGain>bestGain) { |
---|
1770 | bestGain = currGain; |
---|
1771 | bestSplitString = tempStr; |
---|
1772 | for (int jj = 0; jj < 2; jj++) { |
---|
1773 | //dist[jj] = new double[currDist[jj].length]; |
---|
1774 | System.arraycopy(tempDist[jj], 0, dist[jj], 0, |
---|
1775 | dist[jj].length); |
---|
1776 | } |
---|
1777 | } |
---|
1778 | } |
---|
1779 | } |
---|
1780 | |
---|
1781 | // Compute weights |
---|
1782 | int attIndex = att.index(); |
---|
1783 | props[attIndex] = new double[2]; |
---|
1784 | for (int k = 0; k < 2; k++) { |
---|
1785 | props[attIndex][k] = Utils.sum(dist[k]); |
---|
1786 | } |
---|
1787 | if (!(Utils.sum(props[attIndex]) > 0)) { |
---|
1788 | for (int k = 0; k < props[attIndex].length; k++) { |
---|
1789 | props[attIndex][k] = 1.0 / (double)props[attIndex].length; |
---|
1790 | } |
---|
1791 | } else { |
---|
1792 | Utils.normalize(props[attIndex]); |
---|
1793 | } |
---|
1794 | |
---|
1795 | // Compute subset weights |
---|
1796 | subsetWeights[attIndex] = new double[2]; |
---|
1797 | for (int j = 0; j < 2; j++) { |
---|
1798 | subsetWeights[attIndex][j] += Utils.sum(dist[j]); |
---|
1799 | } |
---|
1800 | |
---|
1801 | // Then, for the attribute values that class frequency is 0, split it into the |
---|
1802 | // most frequent branch |
---|
1803 | for (int j=0; j<empty; j++) { |
---|
1804 | if (props[attIndex][0]>=props[attIndex][1]) { |
---|
1805 | if (bestSplitString=="") bestSplitString = "(" + emptyValues[j] + ")"; |
---|
1806 | else bestSplitString += "|" + "(" + emptyValues[j] + ")"; |
---|
1807 | } |
---|
1808 | } |
---|
1809 | |
---|
1810 | // clean gain |
---|
1811 | gains[attIndex] = Math.rint(bestGain*10000000)/10000000.0; |
---|
1812 | |
---|
1813 | dists[attIndex] = dist; |
---|
1814 | return bestSplitString; |
---|
1815 | } |
---|
1816 | |
---|
1817 | |
---|
1818 | /** |
---|
1819 | * Split data into two subsets and store sorted indices and weights for two |
---|
1820 | * successor nodes. |
---|
1821 | * |
---|
1822 | * @param subsetIndices sorted indecis of instances for each attribute for two successor node |
---|
1823 | * @param subsetWeights weights of instances for each attribute for two successor node |
---|
1824 | * @param att attribute the split based on |
---|
1825 | * @param splitPoint split point the split based on if att is numeric |
---|
1826 | * @param splitStr split subset the split based on if att is nominal |
---|
1827 | * @param sortedIndices sorted indices of the instances to be split |
---|
1828 | * @param weights weights of the instances to bes split |
---|
1829 | * @param data training data |
---|
1830 | * @throws Exception if something goes wrong |
---|
1831 | */ |
---|
1832 | protected void splitData(int[][][] subsetIndices, double[][][] subsetWeights, |
---|
1833 | Attribute att, double splitPoint, String splitStr, int[][] sortedIndices, |
---|
1834 | double[][] weights, Instances data) throws Exception { |
---|
1835 | |
---|
1836 | int j; |
---|
1837 | // For each attribute |
---|
1838 | for (int i = 0; i < data.numAttributes(); i++) { |
---|
1839 | if (i==data.classIndex()) continue; |
---|
1840 | int[] num = new int[2]; |
---|
1841 | for (int k = 0; k < 2; k++) { |
---|
1842 | subsetIndices[k][i] = new int[sortedIndices[i].length]; |
---|
1843 | subsetWeights[k][i] = new double[weights[i].length]; |
---|
1844 | } |
---|
1845 | |
---|
1846 | for (j = 0; j < sortedIndices[i].length; j++) { |
---|
1847 | Instance inst = data.instance(sortedIndices[i][j]); |
---|
1848 | if (inst.isMissing(att)) { |
---|
1849 | // Split instance up |
---|
1850 | for (int k = 0; k < 2; k++) { |
---|
1851 | if (m_Props[k] > 0) { |
---|
1852 | subsetIndices[k][i][num[k]] = sortedIndices[i][j]; |
---|
1853 | subsetWeights[k][i][num[k]] = m_Props[k] * weights[i][j]; |
---|
1854 | num[k]++; |
---|
1855 | } |
---|
1856 | } |
---|
1857 | } else { |
---|
1858 | int subset; |
---|
1859 | if (att.isNumeric()) { |
---|
1860 | subset = (inst.value(att) < splitPoint) ? 0 : 1; |
---|
1861 | } else { // nominal attribute |
---|
1862 | if (splitStr.indexOf |
---|
1863 | ("(" + att.value((int)inst.value(att.index()))+")")!=-1) { |
---|
1864 | subset = 0; |
---|
1865 | } else subset = 1; |
---|
1866 | } |
---|
1867 | subsetIndices[subset][i][num[subset]] = sortedIndices[i][j]; |
---|
1868 | subsetWeights[subset][i][num[subset]] = weights[i][j]; |
---|
1869 | num[subset]++; |
---|
1870 | } |
---|
1871 | } |
---|
1872 | |
---|
1873 | // Trim arrays |
---|
1874 | for (int k = 0; k < 2; k++) { |
---|
1875 | int[] copy = new int[num[k]]; |
---|
1876 | System.arraycopy(subsetIndices[k][i], 0, copy, 0, num[k]); |
---|
1877 | subsetIndices[k][i] = copy; |
---|
1878 | double[] copyWeights = new double[num[k]]; |
---|
1879 | System.arraycopy(subsetWeights[k][i], 0 ,copyWeights, 0, num[k]); |
---|
1880 | subsetWeights[k][i] = copyWeights; |
---|
1881 | } |
---|
1882 | } |
---|
1883 | } |
---|
1884 | |
---|
1885 | |
---|
1886 | /** |
---|
1887 | * Compute and return gini gain for given distributions of a node and its |
---|
1888 | * successor nodes. |
---|
1889 | * |
---|
1890 | * @param parentDist class distributions of parent node |
---|
1891 | * @param childDist class distributions of successor nodes |
---|
1892 | * @return Gini gain computed |
---|
1893 | */ |
---|
1894 | protected double computeGiniGain(double[] parentDist, double[][] childDist) { |
---|
1895 | double totalWeight = Utils.sum(parentDist); |
---|
1896 | if (totalWeight==0) return 0; |
---|
1897 | |
---|
1898 | double leftWeight = Utils.sum(childDist[0]); |
---|
1899 | double rightWeight = Utils.sum(childDist[1]); |
---|
1900 | |
---|
1901 | double parentGini = computeGini(parentDist, totalWeight); |
---|
1902 | double leftGini = computeGini(childDist[0],leftWeight); |
---|
1903 | double rightGini = computeGini(childDist[1], rightWeight); |
---|
1904 | |
---|
1905 | return parentGini - leftWeight/totalWeight*leftGini - |
---|
1906 | rightWeight/totalWeight*rightGini; |
---|
1907 | } |
---|
1908 | |
---|
1909 | /** |
---|
1910 | * Compute and return gini index for a given distribution of a node. |
---|
1911 | * |
---|
1912 | * @param dist class distributions |
---|
1913 | * @param total class distributions |
---|
1914 | * @return Gini index of the class distributions |
---|
1915 | */ |
---|
1916 | protected double computeGini(double[] dist, double total) { |
---|
1917 | if (total==0) return 0; |
---|
1918 | double val = 0; |
---|
1919 | for (int i=0; i<dist.length; i++) { |
---|
1920 | val += (dist[i]/total)*(dist[i]/total); |
---|
1921 | } |
---|
1922 | return 1- val; |
---|
1923 | } |
---|
1924 | |
---|
1925 | /** |
---|
1926 | * Compute and return information gain for given distributions of a node |
---|
1927 | * and its successor nodes. |
---|
1928 | * |
---|
1929 | * @param parentDist class distributions of parent node |
---|
1930 | * @param childDist class distributions of successor nodes |
---|
1931 | * @return information gain computed |
---|
1932 | */ |
---|
1933 | protected double computeInfoGain(double[] parentDist, double[][] childDist) { |
---|
1934 | double totalWeight = Utils.sum(parentDist); |
---|
1935 | if (totalWeight==0) return 0; |
---|
1936 | |
---|
1937 | double leftWeight = Utils.sum(childDist[0]); |
---|
1938 | double rightWeight = Utils.sum(childDist[1]); |
---|
1939 | |
---|
1940 | double parentInfo = computeEntropy(parentDist, totalWeight); |
---|
1941 | double leftInfo = computeEntropy(childDist[0],leftWeight); |
---|
1942 | double rightInfo = computeEntropy(childDist[1], rightWeight); |
---|
1943 | |
---|
1944 | return parentInfo - leftWeight/totalWeight*leftInfo - |
---|
1945 | rightWeight/totalWeight*rightInfo; |
---|
1946 | } |
---|
1947 | |
---|
1948 | /** |
---|
1949 | * Compute and return entropy for a given distribution of a node. |
---|
1950 | * |
---|
1951 | * @param dist class distributions |
---|
1952 | * @param total class distributions |
---|
1953 | * @return entropy of the class distributions |
---|
1954 | */ |
---|
1955 | protected double computeEntropy(double[] dist, double total) { |
---|
1956 | if (total==0) return 0; |
---|
1957 | double entropy = 0; |
---|
1958 | for (int i=0; i<dist.length; i++) { |
---|
1959 | if (dist[i]!=0) entropy -= dist[i]/total * Utils.log2(dist[i]/total); |
---|
1960 | } |
---|
1961 | return entropy; |
---|
1962 | } |
---|
1963 | |
---|
1964 | /** |
---|
1965 | * Make the node leaf node. |
---|
1966 | * |
---|
1967 | * @param data training data |
---|
1968 | */ |
---|
1969 | protected void makeLeaf(Instances data) { |
---|
1970 | m_Attribute = null; |
---|
1971 | m_isLeaf = true; |
---|
1972 | m_ClassValue=Utils.maxIndex(m_ClassProbs); |
---|
1973 | m_ClassAttribute = data.classAttribute(); |
---|
1974 | } |
---|
1975 | |
---|
1976 | /** |
---|
1977 | * Computes class probabilities for instance using the decision tree. |
---|
1978 | * |
---|
1979 | * @param instance the instance for which class probabilities is to be computed |
---|
1980 | * @return the class probabilities for the given instance |
---|
1981 | * @throws Exception if something goes wrong |
---|
1982 | */ |
---|
1983 | public double[] distributionForInstance(Instance instance) |
---|
1984 | throws Exception { |
---|
1985 | if (!m_isLeaf) { |
---|
1986 | // value of split attribute is missing |
---|
1987 | if (instance.isMissing(m_Attribute)) { |
---|
1988 | double[] returnedDist = new double[m_ClassProbs.length]; |
---|
1989 | |
---|
1990 | for (int i = 0; i < m_Successors.length; i++) { |
---|
1991 | double[] help = |
---|
1992 | m_Successors[i].distributionForInstance(instance); |
---|
1993 | if (help != null) { |
---|
1994 | for (int j = 0; j < help.length; j++) { |
---|
1995 | returnedDist[j] += m_Props[i] * help[j]; |
---|
1996 | } |
---|
1997 | } |
---|
1998 | } |
---|
1999 | return returnedDist; |
---|
2000 | } |
---|
2001 | |
---|
2002 | // split attribute is nonimal |
---|
2003 | else if (m_Attribute.isNominal()) { |
---|
2004 | if (m_SplitString.indexOf("(" + |
---|
2005 | m_Attribute.value((int)instance.value(m_Attribute)) + ")")!=-1) |
---|
2006 | return m_Successors[0].distributionForInstance(instance); |
---|
2007 | else return m_Successors[1].distributionForInstance(instance); |
---|
2008 | } |
---|
2009 | |
---|
2010 | // split attribute is numeric |
---|
2011 | else { |
---|
2012 | if (instance.value(m_Attribute) < m_SplitValue) |
---|
2013 | return m_Successors[0].distributionForInstance(instance); |
---|
2014 | else |
---|
2015 | return m_Successors[1].distributionForInstance(instance); |
---|
2016 | } |
---|
2017 | } |
---|
2018 | |
---|
2019 | // leaf node |
---|
2020 | else return m_ClassProbs; |
---|
2021 | } |
---|
2022 | |
---|
2023 | /** |
---|
2024 | * Prints the decision tree using the protected toString method from below. |
---|
2025 | * |
---|
2026 | * @return a textual description of the classifier |
---|
2027 | */ |
---|
2028 | public String toString() { |
---|
2029 | if ((m_Distribution == null) && (m_Successors == null)) { |
---|
2030 | return "Best-First: No model built yet."; |
---|
2031 | } |
---|
2032 | return "Best-First Decision Tree\n" + toString(0)+"\n\n" |
---|
2033 | +"Size of the Tree: "+numNodes()+"\n\n" |
---|
2034 | +"Number of Leaf Nodes: "+numLeaves(); |
---|
2035 | } |
---|
2036 | |
---|
2037 | /** |
---|
2038 | * Outputs a tree at a certain level. |
---|
2039 | * |
---|
2040 | * @param level the level at which the tree is to be printed |
---|
2041 | * @return a tree at a certain level. |
---|
2042 | */ |
---|
2043 | protected String toString(int level) { |
---|
2044 | StringBuffer text = new StringBuffer(); |
---|
2045 | // if leaf nodes |
---|
2046 | if (m_Attribute == null) { |
---|
2047 | if (Utils.isMissingValue(m_ClassValue)) { |
---|
2048 | text.append(": null"); |
---|
2049 | } else { |
---|
2050 | double correctNum = Math.rint(m_Distribution[Utils.maxIndex(m_Distribution)]*100)/ |
---|
2051 | 100.0; |
---|
2052 | double wrongNum = Math.rint((Utils.sum(m_Distribution) - |
---|
2053 | m_Distribution[Utils.maxIndex(m_Distribution)])*100)/100.0; |
---|
2054 | String str = "(" + correctNum + "/" + wrongNum + ")"; |
---|
2055 | text.append(": " + m_ClassAttribute.value((int) m_ClassValue)+ str); |
---|
2056 | } |
---|
2057 | } else { |
---|
2058 | for (int j = 0; j < 2; j++) { |
---|
2059 | text.append("\n"); |
---|
2060 | for (int i = 0; i < level; i++) { |
---|
2061 | text.append("| "); |
---|
2062 | } |
---|
2063 | if (j==0) { |
---|
2064 | if (m_Attribute.isNumeric()) |
---|
2065 | text.append(m_Attribute.name() + " < " + m_SplitValue); |
---|
2066 | else |
---|
2067 | text.append(m_Attribute.name() + "=" + m_SplitString); |
---|
2068 | } else { |
---|
2069 | if (m_Attribute.isNumeric()) |
---|
2070 | text.append(m_Attribute.name() + " >= " + m_SplitValue); |
---|
2071 | else |
---|
2072 | text.append(m_Attribute.name() + "!=" + m_SplitString); |
---|
2073 | } |
---|
2074 | text.append(m_Successors[j].toString(level + 1)); |
---|
2075 | } |
---|
2076 | } |
---|
2077 | return text.toString(); |
---|
2078 | } |
---|
2079 | |
---|
2080 | /** |
---|
2081 | * Compute size of the tree. |
---|
2082 | * |
---|
2083 | * @return size of the tree |
---|
2084 | */ |
---|
2085 | public int numNodes() { |
---|
2086 | if (m_isLeaf) { |
---|
2087 | return 1; |
---|
2088 | } else { |
---|
2089 | int size =1; |
---|
2090 | for (int i=0;i<m_Successors.length;i++) { |
---|
2091 | size+=m_Successors[i].numNodes(); |
---|
2092 | } |
---|
2093 | return size; |
---|
2094 | } |
---|
2095 | } |
---|
2096 | |
---|
2097 | /** |
---|
2098 | * Compute number of leaf nodes. |
---|
2099 | * |
---|
2100 | * @return number of leaf nodes |
---|
2101 | */ |
---|
2102 | public int numLeaves() { |
---|
2103 | if (m_isLeaf) return 1; |
---|
2104 | else { |
---|
2105 | int size=0; |
---|
2106 | for (int i=0;i<m_Successors.length;i++) { |
---|
2107 | size+=m_Successors[i].numLeaves(); |
---|
2108 | } |
---|
2109 | return size; |
---|
2110 | } |
---|
2111 | } |
---|
2112 | |
---|
2113 | /** |
---|
2114 | * Returns an enumeration describing the available options. |
---|
2115 | * |
---|
2116 | * @return an enumeration describing the available options. |
---|
2117 | */ |
---|
2118 | public Enumeration listOptions() { |
---|
2119 | Vector result; |
---|
2120 | Enumeration en; |
---|
2121 | |
---|
2122 | result = new Vector(); |
---|
2123 | |
---|
2124 | en = super.listOptions(); |
---|
2125 | while (en.hasMoreElements()) |
---|
2126 | result.addElement(en.nextElement()); |
---|
2127 | |
---|
2128 | result.addElement(new Option( |
---|
2129 | "\tThe pruning strategy.\n" |
---|
2130 | + "\t(default: " + new SelectedTag(PRUNING_POSTPRUNING, TAGS_PRUNING) + ")", |
---|
2131 | "P", 1, "-P " + Tag.toOptionList(TAGS_PRUNING))); |
---|
2132 | |
---|
2133 | result.addElement(new Option( |
---|
2134 | "\tThe minimal number of instances at the terminal nodes.\n" |
---|
2135 | + "\t(default 2)", |
---|
2136 | "M", 1, "-M <min no>")); |
---|
2137 | |
---|
2138 | result.addElement(new Option( |
---|
2139 | "\tThe number of folds used in the pruning.\n" |
---|
2140 | + "\t(default 5)", |
---|
2141 | "N", 5, "-N <num folds>")); |
---|
2142 | |
---|
2143 | result.addElement(new Option( |
---|
2144 | "\tDon't use heuristic search for nominal attributes in multi-class\n" |
---|
2145 | + "\tproblem (default yes).\n", |
---|
2146 | "H", 0, "-H")); |
---|
2147 | |
---|
2148 | result.addElement(new Option( |
---|
2149 | "\tDon't use Gini index for splitting (default yes),\n" |
---|
2150 | + "\tif not information is used.", |
---|
2151 | "G", 0, "-G")); |
---|
2152 | |
---|
2153 | result.addElement(new Option( |
---|
2154 | "\tDon't use error rate in internal cross-validation (default yes), \n" |
---|
2155 | + "\tbut root mean squared error.", |
---|
2156 | "R", 0, "-R")); |
---|
2157 | |
---|
2158 | result.addElement(new Option( |
---|
2159 | "\tUse the 1 SE rule to make pruning decision.\n" |
---|
2160 | + "\t(default no).", |
---|
2161 | "A", 0, "-A")); |
---|
2162 | |
---|
2163 | result.addElement(new Option( |
---|
2164 | "\tPercentage of training data size (0-1]\n" |
---|
2165 | + "\t(default 1).", |
---|
2166 | "C", 0, "-C")); |
---|
2167 | |
---|
2168 | return result.elements(); |
---|
2169 | } |
---|
2170 | |
---|
2171 | /** |
---|
2172 | * Parses the options for this object. <p/> |
---|
2173 | * |
---|
2174 | <!-- options-start --> |
---|
2175 | * Valid options are: <p/> |
---|
2176 | * |
---|
2177 | * <pre> -S <num> |
---|
2178 | * Random number seed. |
---|
2179 | * (default 1)</pre> |
---|
2180 | * |
---|
2181 | * <pre> -D |
---|
2182 | * If set, classifier is run in debug mode and |
---|
2183 | * may output additional info to the console</pre> |
---|
2184 | * |
---|
2185 | * <pre> -P <UNPRUNED|POSTPRUNED|PREPRUNED> |
---|
2186 | * The pruning strategy. |
---|
2187 | * (default: POSTPRUNED)</pre> |
---|
2188 | * |
---|
2189 | * <pre> -M <min no> |
---|
2190 | * The minimal number of instances at the terminal nodes. |
---|
2191 | * (default 2)</pre> |
---|
2192 | * |
---|
2193 | * <pre> -N <num folds> |
---|
2194 | * The number of folds used in the pruning. |
---|
2195 | * (default 5)</pre> |
---|
2196 | * |
---|
2197 | * <pre> -H |
---|
2198 | * Don't use heuristic search for nominal attributes in multi-class |
---|
2199 | * problem (default yes). |
---|
2200 | * </pre> |
---|
2201 | * |
---|
2202 | * <pre> -G |
---|
2203 | * Don't use Gini index for splitting (default yes), |
---|
2204 | * if not information is used.</pre> |
---|
2205 | * |
---|
2206 | * <pre> -R |
---|
2207 | * Don't use error rate in internal cross-validation (default yes), |
---|
2208 | * but root mean squared error.</pre> |
---|
2209 | * |
---|
2210 | * <pre> -A |
---|
2211 | * Use the 1 SE rule to make pruning decision. |
---|
2212 | * (default no).</pre> |
---|
2213 | * |
---|
2214 | * <pre> -C |
---|
2215 | * Percentage of training data size (0-1] |
---|
2216 | * (default 1).</pre> |
---|
2217 | * |
---|
2218 | <!-- options-end --> |
---|
2219 | * |
---|
2220 | * @param options the options to use |
---|
2221 | * @throws Exception if setting of options fails |
---|
2222 | */ |
---|
2223 | public void setOptions(String[] options) throws Exception { |
---|
2224 | String tmpStr; |
---|
2225 | |
---|
2226 | super.setOptions(options); |
---|
2227 | |
---|
2228 | tmpStr = Utils.getOption('M', options); |
---|
2229 | if (tmpStr.length() != 0) |
---|
2230 | setMinNumObj(Integer.parseInt(tmpStr)); |
---|
2231 | else |
---|
2232 | setMinNumObj(2); |
---|
2233 | |
---|
2234 | tmpStr = Utils.getOption('N', options); |
---|
2235 | if (tmpStr.length() != 0) |
---|
2236 | setNumFoldsPruning(Integer.parseInt(tmpStr)); |
---|
2237 | else |
---|
2238 | setNumFoldsPruning(5); |
---|
2239 | |
---|
2240 | tmpStr = Utils.getOption('C', options); |
---|
2241 | if (tmpStr.length()!=0) |
---|
2242 | setSizePer(Double.parseDouble(tmpStr)); |
---|
2243 | else |
---|
2244 | setSizePer(1); |
---|
2245 | |
---|
2246 | tmpStr = Utils.getOption('P', options); |
---|
2247 | if (tmpStr.length() != 0) |
---|
2248 | setPruningStrategy(new SelectedTag(tmpStr, TAGS_PRUNING)); |
---|
2249 | else |
---|
2250 | setPruningStrategy(new SelectedTag(PRUNING_POSTPRUNING, TAGS_PRUNING)); |
---|
2251 | |
---|
2252 | setHeuristic(!Utils.getFlag('H',options)); |
---|
2253 | |
---|
2254 | setUseGini(!Utils.getFlag('G',options)); |
---|
2255 | |
---|
2256 | setUseErrorRate(!Utils.getFlag('R',options)); |
---|
2257 | |
---|
2258 | setUseOneSE(Utils.getFlag('A',options)); |
---|
2259 | } |
---|
2260 | |
---|
2261 | /** |
---|
2262 | * Gets the current settings of the Classifier. |
---|
2263 | * |
---|
2264 | * @return the current settings of the Classifier |
---|
2265 | */ |
---|
2266 | public String[] getOptions() { |
---|
2267 | int i; |
---|
2268 | Vector result; |
---|
2269 | String[] options; |
---|
2270 | |
---|
2271 | result = new Vector(); |
---|
2272 | |
---|
2273 | options = super.getOptions(); |
---|
2274 | for (i = 0; i < options.length; i++) |
---|
2275 | result.add(options[i]); |
---|
2276 | |
---|
2277 | result.add("-M"); |
---|
2278 | result.add("" + getMinNumObj()); |
---|
2279 | |
---|
2280 | result.add("-N"); |
---|
2281 | result.add("" + getNumFoldsPruning()); |
---|
2282 | |
---|
2283 | if (!getHeuristic()) |
---|
2284 | result.add("-H"); |
---|
2285 | |
---|
2286 | if (!getUseGini()) |
---|
2287 | result.add("-G"); |
---|
2288 | |
---|
2289 | if (!getUseErrorRate()) |
---|
2290 | result.add("-R"); |
---|
2291 | |
---|
2292 | if (getUseOneSE()) |
---|
2293 | result.add("-A"); |
---|
2294 | |
---|
2295 | result.add("-C"); |
---|
2296 | result.add("" + getSizePer()); |
---|
2297 | |
---|
2298 | result.add("-P"); |
---|
2299 | result.add("" + getPruningStrategy()); |
---|
2300 | |
---|
2301 | return (String[]) result.toArray(new String[result.size()]); |
---|
2302 | } |
---|
2303 | |
---|
2304 | /** |
---|
2305 | * Return an enumeration of the measure names. |
---|
2306 | * |
---|
2307 | * @return an enumeration of the measure names |
---|
2308 | */ |
---|
2309 | public Enumeration enumerateMeasures() { |
---|
2310 | Vector result = new Vector(); |
---|
2311 | |
---|
2312 | result.addElement("measureTreeSize"); |
---|
2313 | |
---|
2314 | return result.elements(); |
---|
2315 | } |
---|
2316 | |
---|
2317 | /** |
---|
2318 | * Return number of tree size. |
---|
2319 | * |
---|
2320 | * @return number of tree size |
---|
2321 | */ |
---|
2322 | public double measureTreeSize() { |
---|
2323 | return numNodes(); |
---|
2324 | } |
---|
2325 | |
---|
2326 | /** |
---|
2327 | * Returns the value of the named measure |
---|
2328 | * |
---|
2329 | * @param additionalMeasureName the name of the measure to query for its value |
---|
2330 | * @return the value of the named measure |
---|
2331 | * @throws IllegalArgumentException if the named measure is not supported |
---|
2332 | */ |
---|
2333 | public double getMeasure(String additionalMeasureName) { |
---|
2334 | if (additionalMeasureName.compareToIgnoreCase("measureTreeSize") == 0) { |
---|
2335 | return measureTreeSize(); |
---|
2336 | } else { |
---|
2337 | throw new IllegalArgumentException(additionalMeasureName |
---|
2338 | + " not supported (Best-First)"); |
---|
2339 | } |
---|
2340 | } |
---|
2341 | |
---|
2342 | /** |
---|
2343 | * Returns the tip text for this property |
---|
2344 | * |
---|
2345 | * @return tip text for this property suitable for |
---|
2346 | * displaying in the explorer/experimenter gui |
---|
2347 | */ |
---|
2348 | public String pruningStrategyTipText() { |
---|
2349 | return "Sets the pruning strategy."; |
---|
2350 | } |
---|
2351 | |
---|
2352 | /** |
---|
2353 | * Sets the pruning strategy. |
---|
2354 | * |
---|
2355 | * @param value the strategy |
---|
2356 | */ |
---|
2357 | public void setPruningStrategy(SelectedTag value) { |
---|
2358 | if (value.getTags() == TAGS_PRUNING) { |
---|
2359 | m_PruningStrategy = value.getSelectedTag().getID(); |
---|
2360 | } |
---|
2361 | } |
---|
2362 | |
---|
2363 | /** |
---|
2364 | * Gets the pruning strategy. |
---|
2365 | * |
---|
2366 | * @return the current strategy. |
---|
2367 | */ |
---|
2368 | public SelectedTag getPruningStrategy() { |
---|
2369 | return new SelectedTag(m_PruningStrategy, TAGS_PRUNING); |
---|
2370 | } |
---|
2371 | |
---|
2372 | /** |
---|
2373 | * Returns the tip text for this property |
---|
2374 | * |
---|
2375 | * @return tip text for this property suitable for |
---|
2376 | * displaying in the explorer/experimenter gui |
---|
2377 | */ |
---|
2378 | public String minNumObjTipText() { |
---|
2379 | return "Set minimal number of instances at the terminal nodes."; |
---|
2380 | } |
---|
2381 | |
---|
2382 | /** |
---|
2383 | * Set minimal number of instances at the terminal nodes. |
---|
2384 | * |
---|
2385 | * @param value minimal number of instances at the terminal nodes |
---|
2386 | */ |
---|
2387 | public void setMinNumObj(int value) { |
---|
2388 | m_minNumObj = value; |
---|
2389 | } |
---|
2390 | |
---|
2391 | /** |
---|
2392 | * Get minimal number of instances at the terminal nodes. |
---|
2393 | * |
---|
2394 | * @return minimal number of instances at the terminal nodes |
---|
2395 | */ |
---|
2396 | public int getMinNumObj() { |
---|
2397 | return m_minNumObj; |
---|
2398 | } |
---|
2399 | |
---|
2400 | /** |
---|
2401 | * Returns the tip text for this property |
---|
2402 | * |
---|
2403 | * @return tip text for this property suitable for |
---|
2404 | * displaying in the explorer/experimenter gui |
---|
2405 | */ |
---|
2406 | public String numFoldsPruningTipText() { |
---|
2407 | return "Number of folds in internal cross-validation."; |
---|
2408 | } |
---|
2409 | |
---|
2410 | /** |
---|
2411 | * Set number of folds in internal cross-validation. |
---|
2412 | * |
---|
2413 | * @param value the number of folds |
---|
2414 | */ |
---|
2415 | public void setNumFoldsPruning(int value) { |
---|
2416 | m_numFoldsPruning = value; |
---|
2417 | } |
---|
2418 | |
---|
2419 | /** |
---|
2420 | * Set number of folds in internal cross-validation. |
---|
2421 | * |
---|
2422 | * @return number of folds in internal cross-validation |
---|
2423 | */ |
---|
2424 | public int getNumFoldsPruning() { |
---|
2425 | return m_numFoldsPruning; |
---|
2426 | } |
---|
2427 | |
---|
2428 | /** |
---|
2429 | * Returns the tip text for this property |
---|
2430 | * |
---|
2431 | * @return tip text for this property suitable for |
---|
2432 | * displaying in the explorer/experimenter gui. |
---|
2433 | */ |
---|
2434 | public String heuristicTipText() { |
---|
2435 | return "If heuristic search is used for binary split for nominal attributes."; |
---|
2436 | } |
---|
2437 | |
---|
2438 | /** |
---|
2439 | * Set if use heuristic search for nominal attributes in multi-class problems. |
---|
2440 | * |
---|
2441 | * @param value if use heuristic search for nominal attributes in |
---|
2442 | * multi-class problems |
---|
2443 | */ |
---|
2444 | public void setHeuristic(boolean value) { |
---|
2445 | m_Heuristic = value; |
---|
2446 | } |
---|
2447 | |
---|
2448 | /** |
---|
2449 | * Get if use heuristic search for nominal attributes in multi-class problems. |
---|
2450 | * |
---|
2451 | * @return if use heuristic search for nominal attributes in |
---|
2452 | * multi-class problems |
---|
2453 | */ |
---|
2454 | public boolean getHeuristic() { |
---|
2455 | return m_Heuristic; |
---|
2456 | } |
---|
2457 | |
---|
2458 | /** |
---|
2459 | * Returns the tip text for this property |
---|
2460 | * |
---|
2461 | * @return tip text for this property suitable for |
---|
2462 | * displaying in the explorer/experimenter gui. |
---|
2463 | */ |
---|
2464 | public String useGiniTipText() { |
---|
2465 | return "If true the Gini index is used for splitting criterion, otherwise the information is used."; |
---|
2466 | } |
---|
2467 | |
---|
2468 | /** |
---|
2469 | * Set if use Gini index as splitting criterion. |
---|
2470 | * |
---|
2471 | * @param value if use Gini index splitting criterion |
---|
2472 | */ |
---|
2473 | public void setUseGini(boolean value) { |
---|
2474 | m_UseGini = value; |
---|
2475 | } |
---|
2476 | |
---|
2477 | /** |
---|
2478 | * Get if use Gini index as splitting criterion. |
---|
2479 | * |
---|
2480 | * @return if use Gini index as splitting criterion |
---|
2481 | */ |
---|
2482 | public boolean getUseGini() { |
---|
2483 | return m_UseGini; |
---|
2484 | } |
---|
2485 | |
---|
2486 | /** |
---|
2487 | * Returns the tip text for this property |
---|
2488 | * |
---|
2489 | * @return tip text for this property suitable for |
---|
2490 | * displaying in the explorer/experimenter gui. |
---|
2491 | */ |
---|
2492 | public String useErrorRateTipText() { |
---|
2493 | return "If error rate is used as error estimate. if not, root mean squared error is used."; |
---|
2494 | } |
---|
2495 | |
---|
2496 | /** |
---|
2497 | * Set if use error rate in internal cross-validation. |
---|
2498 | * |
---|
2499 | * @param value if use error rate in internal cross-validation |
---|
2500 | */ |
---|
2501 | public void setUseErrorRate(boolean value) { |
---|
2502 | m_UseErrorRate = value; |
---|
2503 | } |
---|
2504 | |
---|
2505 | /** |
---|
2506 | * Get if use error rate in internal cross-validation. |
---|
2507 | * |
---|
2508 | * @return if use error rate in internal cross-validation. |
---|
2509 | */ |
---|
2510 | public boolean getUseErrorRate() { |
---|
2511 | return m_UseErrorRate; |
---|
2512 | } |
---|
2513 | |
---|
2514 | /** |
---|
2515 | * Returns the tip text for this property |
---|
2516 | * |
---|
2517 | * @return tip text for this property suitable for |
---|
2518 | * displaying in the explorer/experimenter gui. |
---|
2519 | */ |
---|
2520 | public String useOneSETipText() { |
---|
2521 | return "Use the 1SE rule to make pruning decision."; |
---|
2522 | } |
---|
2523 | |
---|
2524 | /** |
---|
2525 | * Set if use the 1SE rule to choose final model. |
---|
2526 | * |
---|
2527 | * @param value if use the 1SE rule to choose final model |
---|
2528 | */ |
---|
2529 | public void setUseOneSE(boolean value) { |
---|
2530 | m_UseOneSE = value; |
---|
2531 | } |
---|
2532 | |
---|
2533 | /** |
---|
2534 | * Get if use the 1SE rule to choose final model. |
---|
2535 | * |
---|
2536 | * @return if use the 1SE rule to choose final model |
---|
2537 | */ |
---|
2538 | public boolean getUseOneSE() { |
---|
2539 | return m_UseOneSE; |
---|
2540 | } |
---|
2541 | |
---|
2542 | /** |
---|
2543 | * Returns the tip text for this property |
---|
2544 | * |
---|
2545 | * @return tip text for this property suitable for |
---|
2546 | * displaying in the explorer/experimenter gui. |
---|
2547 | */ |
---|
2548 | public String sizePerTipText() { |
---|
2549 | return "The percentage of the training set size (0-1, 0 not included)."; |
---|
2550 | } |
---|
2551 | |
---|
2552 | /** |
---|
2553 | * Set training set size. |
---|
2554 | * |
---|
2555 | * @param value training set size |
---|
2556 | */ |
---|
2557 | public void setSizePer(double value) { |
---|
2558 | if ((value <= 0) || (value > 1)) |
---|
2559 | System.err.println( |
---|
2560 | "The percentage of the training set size must be in range 0 to 1 " |
---|
2561 | + "(0 not included) - ignored!"); |
---|
2562 | else |
---|
2563 | m_SizePer = value; |
---|
2564 | } |
---|
2565 | |
---|
2566 | /** |
---|
2567 | * Get training set size. |
---|
2568 | * |
---|
2569 | * @return training set size |
---|
2570 | */ |
---|
2571 | public double getSizePer() { |
---|
2572 | return m_SizePer; |
---|
2573 | } |
---|
2574 | |
---|
2575 | /** |
---|
2576 | * Returns the revision string. |
---|
2577 | * |
---|
2578 | * @return the revision |
---|
2579 | */ |
---|
2580 | public String getRevision() { |
---|
2581 | return RevisionUtils.extract("$Revision: 5987 $"); |
---|
2582 | } |
---|
2583 | |
---|
2584 | /** |
---|
2585 | * Main method. |
---|
2586 | * |
---|
2587 | * @param args the options for the classifier |
---|
2588 | */ |
---|
2589 | public static void main(String[] args) { |
---|
2590 | runClassifier(new BFTree(), args); |
---|
2591 | } |
---|
2592 | } |
---|