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 | * MILR.java |
---|
19 | * Copyright (C) 2005 University of Waikato, Hamilton, New Zealand |
---|
20 | * |
---|
21 | */ |
---|
22 | |
---|
23 | package weka.classifiers.mi; |
---|
24 | |
---|
25 | import weka.classifiers.Classifier; |
---|
26 | import weka.classifiers.AbstractClassifier; |
---|
27 | import weka.core.Capabilities; |
---|
28 | import weka.core.Instance; |
---|
29 | import weka.core.Instances; |
---|
30 | import weka.core.MultiInstanceCapabilitiesHandler; |
---|
31 | import weka.core.Optimization; |
---|
32 | import weka.core.Option; |
---|
33 | import weka.core.OptionHandler; |
---|
34 | import weka.core.RevisionUtils; |
---|
35 | import weka.core.SelectedTag; |
---|
36 | import weka.core.Tag; |
---|
37 | import weka.core.Utils; |
---|
38 | import weka.core.Capabilities.Capability; |
---|
39 | |
---|
40 | import java.util.Enumeration; |
---|
41 | import java.util.Vector; |
---|
42 | |
---|
43 | /** |
---|
44 | <!-- globalinfo-start --> |
---|
45 | * Uses either standard or collective multi-instance assumption, but within linear regression. For the collective assumption, it offers arithmetic or geometric mean for the posteriors. |
---|
46 | * <p/> |
---|
47 | <!-- globalinfo-end --> |
---|
48 | * |
---|
49 | <!-- options-start --> |
---|
50 | * Valid options are: <p/> |
---|
51 | * |
---|
52 | * <pre> -D |
---|
53 | * Turn on debugging output.</pre> |
---|
54 | * |
---|
55 | * <pre> -R <ridge> |
---|
56 | * Set the ridge in the log-likelihood.</pre> |
---|
57 | * |
---|
58 | * <pre> -A [0|1|2] |
---|
59 | * Defines the type of algorithm: |
---|
60 | * 0. standard MI assumption |
---|
61 | * 1. collective MI assumption, arithmetic mean for posteriors |
---|
62 | * 2. collective MI assumption, geometric mean for posteriors</pre> |
---|
63 | * |
---|
64 | <!-- options-end --> |
---|
65 | * |
---|
66 | * @author Eibe Frank (eibe@cs.waikato.ac.nz) |
---|
67 | * @author Xin Xu (xx5@cs.waikato.ac.nz) |
---|
68 | * @version $Revision: 5928 $ |
---|
69 | */ |
---|
70 | public class MILR |
---|
71 | extends AbstractClassifier |
---|
72 | implements OptionHandler, MultiInstanceCapabilitiesHandler { |
---|
73 | |
---|
74 | /** for serialization */ |
---|
75 | static final long serialVersionUID = 1996101190172373826L; |
---|
76 | |
---|
77 | protected double[] m_Par; |
---|
78 | |
---|
79 | /** The number of the class labels */ |
---|
80 | protected int m_NumClasses; |
---|
81 | |
---|
82 | /** The ridge parameter. */ |
---|
83 | protected double m_Ridge = 1e-6; |
---|
84 | |
---|
85 | /** Class labels for each bag */ |
---|
86 | protected int[] m_Classes; |
---|
87 | |
---|
88 | /** MI data */ |
---|
89 | protected double[][][] m_Data; |
---|
90 | |
---|
91 | /** All attribute names */ |
---|
92 | protected Instances m_Attributes; |
---|
93 | |
---|
94 | protected double[] xMean = null, xSD = null; |
---|
95 | |
---|
96 | /** the type of processing */ |
---|
97 | protected int m_AlgorithmType = ALGORITHMTYPE_DEFAULT; |
---|
98 | |
---|
99 | /** standard MI assumption */ |
---|
100 | public static final int ALGORITHMTYPE_DEFAULT = 0; |
---|
101 | /** collective MI assumption, arithmetic mean for posteriors */ |
---|
102 | public static final int ALGORITHMTYPE_ARITHMETIC = 1; |
---|
103 | /** collective MI assumption, geometric mean for posteriors */ |
---|
104 | public static final int ALGORITHMTYPE_GEOMETRIC = 2; |
---|
105 | /** the types of algorithms */ |
---|
106 | public static final Tag [] TAGS_ALGORITHMTYPE = { |
---|
107 | new Tag(ALGORITHMTYPE_DEFAULT, "standard MI assumption"), |
---|
108 | new Tag(ALGORITHMTYPE_ARITHMETIC, "collective MI assumption, arithmetic mean for posteriors"), |
---|
109 | new Tag(ALGORITHMTYPE_GEOMETRIC, "collective MI assumption, geometric mean for posteriors"), |
---|
110 | }; |
---|
111 | |
---|
112 | /** |
---|
113 | * Returns the tip text for this property |
---|
114 | * |
---|
115 | * @return tip text for this property suitable for |
---|
116 | * displaying in the explorer/experimenter gui |
---|
117 | */ |
---|
118 | public String globalInfo() { |
---|
119 | return |
---|
120 | "Uses either standard or collective multi-instance assumption, but " |
---|
121 | + "within linear regression. For the collective assumption, it offers " |
---|
122 | + "arithmetic or geometric mean for the posteriors."; |
---|
123 | } |
---|
124 | |
---|
125 | /** |
---|
126 | * Returns an enumeration describing the available options |
---|
127 | * |
---|
128 | * @return an enumeration of all the available options |
---|
129 | */ |
---|
130 | public Enumeration listOptions() { |
---|
131 | Vector result = new Vector(); |
---|
132 | |
---|
133 | result.addElement(new Option( |
---|
134 | "\tTurn on debugging output.", |
---|
135 | "D", 0, "-D")); |
---|
136 | |
---|
137 | result.addElement(new Option( |
---|
138 | "\tSet the ridge in the log-likelihood.", |
---|
139 | "R", 1, "-R <ridge>")); |
---|
140 | |
---|
141 | result.addElement(new Option( |
---|
142 | "\tDefines the type of algorithm:\n" |
---|
143 | + "\t 0. standard MI assumption\n" |
---|
144 | + "\t 1. collective MI assumption, arithmetic mean for posteriors\n" |
---|
145 | + "\t 2. collective MI assumption, geometric mean for posteriors", |
---|
146 | "A", 1, "-A [0|1|2]")); |
---|
147 | |
---|
148 | return result.elements(); |
---|
149 | } |
---|
150 | |
---|
151 | /** |
---|
152 | * Parses a given list of options. |
---|
153 | * |
---|
154 | * @param options the list of options as an array of strings |
---|
155 | * @throws Exception if an option is not supported |
---|
156 | */ |
---|
157 | public void setOptions(String[] options) throws Exception { |
---|
158 | String tmpStr; |
---|
159 | |
---|
160 | setDebug(Utils.getFlag('D', options)); |
---|
161 | |
---|
162 | tmpStr = Utils.getOption('R', options); |
---|
163 | if (tmpStr.length() != 0) |
---|
164 | setRidge(Double.parseDouble(tmpStr)); |
---|
165 | else |
---|
166 | setRidge(1.0e-6); |
---|
167 | |
---|
168 | tmpStr = Utils.getOption('A', options); |
---|
169 | if (tmpStr.length() != 0) { |
---|
170 | setAlgorithmType(new SelectedTag(Integer.parseInt(tmpStr), TAGS_ALGORITHMTYPE)); |
---|
171 | } else { |
---|
172 | setAlgorithmType(new SelectedTag(ALGORITHMTYPE_DEFAULT, TAGS_ALGORITHMTYPE)); |
---|
173 | } |
---|
174 | } |
---|
175 | |
---|
176 | /** |
---|
177 | * Gets the current settings of the classifier. |
---|
178 | * |
---|
179 | * @return an array of strings suitable for passing to setOptions |
---|
180 | */ |
---|
181 | public String[] getOptions() { |
---|
182 | Vector result; |
---|
183 | |
---|
184 | result = new Vector(); |
---|
185 | |
---|
186 | if (getDebug()) |
---|
187 | result.add("-D"); |
---|
188 | |
---|
189 | result.add("-R"); |
---|
190 | result.add("" + getRidge()); |
---|
191 | |
---|
192 | result.add("-A"); |
---|
193 | result.add("" + m_AlgorithmType); |
---|
194 | |
---|
195 | return (String[]) result.toArray(new String[result.size()]); |
---|
196 | } |
---|
197 | |
---|
198 | /** |
---|
199 | * Returns the tip text for this property |
---|
200 | * |
---|
201 | * @return tip text for this property suitable for |
---|
202 | * displaying in the explorer/experimenter gui |
---|
203 | */ |
---|
204 | public String ridgeTipText() { |
---|
205 | return "The ridge in the log-likelihood."; |
---|
206 | } |
---|
207 | |
---|
208 | /** |
---|
209 | * Sets the ridge in the log-likelihood. |
---|
210 | * |
---|
211 | * @param ridge the ridge |
---|
212 | */ |
---|
213 | public void setRidge(double ridge) { |
---|
214 | m_Ridge = ridge; |
---|
215 | } |
---|
216 | |
---|
217 | /** |
---|
218 | * Gets the ridge in the log-likelihood. |
---|
219 | * |
---|
220 | * @return the ridge |
---|
221 | */ |
---|
222 | public double getRidge() { |
---|
223 | return m_Ridge; |
---|
224 | } |
---|
225 | |
---|
226 | /** |
---|
227 | * Returns the tip text for this property |
---|
228 | * |
---|
229 | * @return tip text for this property suitable for |
---|
230 | * displaying in the explorer/experimenter gui |
---|
231 | */ |
---|
232 | public String algorithmTypeTipText() { |
---|
233 | return "The mean type for the posteriors."; |
---|
234 | } |
---|
235 | |
---|
236 | /** |
---|
237 | * Gets the type of algorithm. |
---|
238 | * |
---|
239 | * @return the algorithm type |
---|
240 | */ |
---|
241 | public SelectedTag getAlgorithmType() { |
---|
242 | return new SelectedTag(m_AlgorithmType, TAGS_ALGORITHMTYPE); |
---|
243 | } |
---|
244 | |
---|
245 | /** |
---|
246 | * Sets the algorithm type. |
---|
247 | * |
---|
248 | * @param newType the new algorithm type |
---|
249 | */ |
---|
250 | public void setAlgorithmType(SelectedTag newType) { |
---|
251 | if (newType.getTags() == TAGS_ALGORITHMTYPE) { |
---|
252 | m_AlgorithmType = newType.getSelectedTag().getID(); |
---|
253 | } |
---|
254 | } |
---|
255 | |
---|
256 | private class OptEng |
---|
257 | extends Optimization { |
---|
258 | |
---|
259 | /** the type to use |
---|
260 | * @see MILR#TAGS_ALGORITHMTYPE */ |
---|
261 | private int m_Type; |
---|
262 | |
---|
263 | /** |
---|
264 | * initializes the object |
---|
265 | * |
---|
266 | * @param type the type top use |
---|
267 | * @see MILR#TAGS_ALGORITHMTYPE |
---|
268 | */ |
---|
269 | public OptEng(int type) { |
---|
270 | super(); |
---|
271 | |
---|
272 | m_Type = type; |
---|
273 | } |
---|
274 | |
---|
275 | /** |
---|
276 | * Evaluate objective function |
---|
277 | * @param x the current values of variables |
---|
278 | * @return the value of the objective function |
---|
279 | */ |
---|
280 | protected double objectiveFunction(double[] x){ |
---|
281 | double nll = 0; // -LogLikelihood |
---|
282 | |
---|
283 | switch (m_Type) { |
---|
284 | case ALGORITHMTYPE_DEFAULT: |
---|
285 | for(int i=0; i<m_Classes.length; i++){ // ith bag |
---|
286 | int nI = m_Data[i][0].length; // numInstances in ith bag |
---|
287 | double bag = 0.0, // NLL of each bag |
---|
288 | prod = 0.0; // Log-prob. |
---|
289 | |
---|
290 | for(int j=0; j<nI; j++){ |
---|
291 | double exp=0.0; |
---|
292 | for(int k=m_Data[i].length-1; k>=0; k--) |
---|
293 | exp += m_Data[i][k][j]*x[k+1]; |
---|
294 | exp += x[0]; |
---|
295 | exp = Math.exp(exp); |
---|
296 | |
---|
297 | if(m_Classes[i]==1) |
---|
298 | prod -= Math.log(1.0+exp); |
---|
299 | else |
---|
300 | bag += Math.log(1.0+exp); |
---|
301 | } |
---|
302 | |
---|
303 | if(m_Classes[i]==1) |
---|
304 | bag = -Math.log(1.0-Math.exp(prod)); |
---|
305 | |
---|
306 | nll += bag; |
---|
307 | } |
---|
308 | break; |
---|
309 | |
---|
310 | case ALGORITHMTYPE_ARITHMETIC: |
---|
311 | for(int i=0; i<m_Classes.length; i++){ // ith bag |
---|
312 | int nI = m_Data[i][0].length; // numInstances in ith bag |
---|
313 | double bag = 0; // NLL of each bag |
---|
314 | |
---|
315 | for(int j=0; j<nI; j++){ |
---|
316 | double exp=0.0; |
---|
317 | for(int k=m_Data[i].length-1; k>=0; k--) |
---|
318 | exp += m_Data[i][k][j]*x[k+1]; |
---|
319 | exp += x[0]; |
---|
320 | exp = Math.exp(exp); |
---|
321 | |
---|
322 | if(m_Classes[i] == 1) |
---|
323 | bag += 1.0-1.0/(1.0+exp); // To avoid exp infinite |
---|
324 | else |
---|
325 | bag += 1.0/(1.0+exp); |
---|
326 | } |
---|
327 | bag /= (double)nI; |
---|
328 | |
---|
329 | nll -= Math.log(bag); |
---|
330 | } |
---|
331 | break; |
---|
332 | |
---|
333 | case ALGORITHMTYPE_GEOMETRIC: |
---|
334 | for(int i=0; i<m_Classes.length; i++){ // ith bag |
---|
335 | int nI = m_Data[i][0].length; // numInstances in ith bag |
---|
336 | double bag = 0; // Log-prob. |
---|
337 | |
---|
338 | for(int j=0; j<nI; j++){ |
---|
339 | double exp=0.0; |
---|
340 | for(int k=m_Data[i].length-1; k>=0; k--) |
---|
341 | exp += m_Data[i][k][j]*x[k+1]; |
---|
342 | exp += x[0]; |
---|
343 | |
---|
344 | if(m_Classes[i]==1) |
---|
345 | bag -= exp/(double)nI; |
---|
346 | else |
---|
347 | bag += exp/(double)nI; |
---|
348 | } |
---|
349 | |
---|
350 | nll += Math.log(1.0+Math.exp(bag)); |
---|
351 | } |
---|
352 | break; |
---|
353 | } |
---|
354 | |
---|
355 | // ridge: note that intercepts NOT included |
---|
356 | for(int r=1; r<x.length; r++) |
---|
357 | nll += m_Ridge*x[r]*x[r]; |
---|
358 | |
---|
359 | return nll; |
---|
360 | } |
---|
361 | |
---|
362 | /** |
---|
363 | * Evaluate Jacobian vector |
---|
364 | * @param x the current values of variables |
---|
365 | * @return the gradient vector |
---|
366 | */ |
---|
367 | protected double[] evaluateGradient(double[] x){ |
---|
368 | double[] grad = new double[x.length]; |
---|
369 | |
---|
370 | switch (m_Type) { |
---|
371 | case ALGORITHMTYPE_DEFAULT: |
---|
372 | for(int i=0; i<m_Classes.length; i++){ // ith bag |
---|
373 | int nI = m_Data[i][0].length; // numInstances in ith bag |
---|
374 | |
---|
375 | double denom = 0.0; // denominator, in log-scale |
---|
376 | double[] bag = new double[grad.length]; //gradient update with ith bag |
---|
377 | |
---|
378 | for(int j=0; j<nI; j++){ |
---|
379 | // Compute exp(b0+b1*Xi1j+...)/[1+exp(b0+b1*Xi1j+...)] |
---|
380 | double exp=0.0; |
---|
381 | for(int k=m_Data[i].length-1; k>=0; k--) |
---|
382 | exp += m_Data[i][k][j]*x[k+1]; |
---|
383 | exp += x[0]; |
---|
384 | exp = Math.exp(exp)/(1.0+Math.exp(exp)); |
---|
385 | |
---|
386 | if(m_Classes[i]==1) |
---|
387 | // Bug fix: it used to be denom += Math.log(1.0+exp); |
---|
388 | // Fixed 21 Jan 2005 (Eibe) |
---|
389 | denom -= Math.log(1.0-exp); |
---|
390 | |
---|
391 | // Instance-wise update of dNLL/dBk |
---|
392 | for(int p=0; p<x.length; p++){ // pth variable |
---|
393 | double m = 1.0; |
---|
394 | if(p>0) m=m_Data[i][p-1][j]; |
---|
395 | bag[p] += m*exp; |
---|
396 | } |
---|
397 | } |
---|
398 | |
---|
399 | denom = Math.exp(denom); |
---|
400 | |
---|
401 | // Bag-wise update of dNLL/dBk |
---|
402 | for(int q=0; q<grad.length; q++){ |
---|
403 | if(m_Classes[i]==1) |
---|
404 | grad[q] -= bag[q]/(denom-1.0); |
---|
405 | else |
---|
406 | grad[q] += bag[q]; |
---|
407 | } |
---|
408 | } |
---|
409 | break; |
---|
410 | |
---|
411 | case ALGORITHMTYPE_ARITHMETIC: |
---|
412 | for(int i=0; i<m_Classes.length; i++){ // ith bag |
---|
413 | int nI = m_Data[i][0].length; // numInstances in ith bag |
---|
414 | |
---|
415 | double denom=0.0; |
---|
416 | double[] numrt = new double[x.length]; |
---|
417 | |
---|
418 | for(int j=0; j<nI; j++){ |
---|
419 | // Compute exp(b0+b1*Xi1j+...)/[1+exp(b0+b1*Xi1j+...)] |
---|
420 | double exp=0.0; |
---|
421 | for(int k=m_Data[i].length-1; k>=0; k--) |
---|
422 | exp += m_Data[i][k][j]*x[k+1]; |
---|
423 | exp += x[0]; |
---|
424 | exp = Math.exp(exp); |
---|
425 | if(m_Classes[i]==1) |
---|
426 | denom += exp/(1.0+exp); |
---|
427 | else |
---|
428 | denom += 1.0/(1.0+exp); |
---|
429 | |
---|
430 | // Instance-wise update of dNLL/dBk |
---|
431 | for(int p=0; p<x.length; p++){ // pth variable |
---|
432 | double m = 1.0; |
---|
433 | if(p>0) m=m_Data[i][p-1][j]; |
---|
434 | numrt[p] += m*exp/((1.0+exp)*(1.0+exp)); |
---|
435 | } |
---|
436 | } |
---|
437 | |
---|
438 | // Bag-wise update of dNLL/dBk |
---|
439 | for(int q=0; q<grad.length; q++){ |
---|
440 | if(m_Classes[i]==1) |
---|
441 | grad[q] -= numrt[q]/denom; |
---|
442 | else |
---|
443 | grad[q] += numrt[q]/denom; |
---|
444 | } |
---|
445 | } |
---|
446 | break; |
---|
447 | |
---|
448 | case ALGORITHMTYPE_GEOMETRIC: |
---|
449 | for(int i=0; i<m_Classes.length; i++){ // ith bag |
---|
450 | int nI = m_Data[i][0].length; // numInstances in ith bag |
---|
451 | double bag = 0; |
---|
452 | double[] sumX = new double[x.length]; |
---|
453 | for(int j=0; j<nI; j++){ |
---|
454 | // Compute exp(b0+b1*Xi1j+...)/[1+exp(b0+b1*Xi1j+...)] |
---|
455 | double exp=0.0; |
---|
456 | for(int k=m_Data[i].length-1; k>=0; k--) |
---|
457 | exp += m_Data[i][k][j]*x[k+1]; |
---|
458 | exp += x[0]; |
---|
459 | |
---|
460 | if(m_Classes[i]==1){ |
---|
461 | bag -= exp/(double)nI; |
---|
462 | for(int q=0; q<grad.length; q++){ |
---|
463 | double m = 1.0; |
---|
464 | if(q>0) m=m_Data[i][q-1][j]; |
---|
465 | sumX[q] -= m/(double)nI; |
---|
466 | } |
---|
467 | } |
---|
468 | else{ |
---|
469 | bag += exp/(double)nI; |
---|
470 | for(int q=0; q<grad.length; q++){ |
---|
471 | double m = 1.0; |
---|
472 | if(q>0) m=m_Data[i][q-1][j]; |
---|
473 | sumX[q] += m/(double)nI; |
---|
474 | } |
---|
475 | } |
---|
476 | } |
---|
477 | |
---|
478 | for(int p=0; p<x.length; p++) |
---|
479 | grad[p] += Math.exp(bag)*sumX[p]/(1.0+Math.exp(bag)); |
---|
480 | } |
---|
481 | break; |
---|
482 | } |
---|
483 | |
---|
484 | // ridge: note that intercepts NOT included |
---|
485 | for(int r=1; r<x.length; r++){ |
---|
486 | grad[r] += 2.0*m_Ridge*x[r]; |
---|
487 | } |
---|
488 | |
---|
489 | return grad; |
---|
490 | } |
---|
491 | |
---|
492 | /** |
---|
493 | * Returns the revision string. |
---|
494 | * |
---|
495 | * @return the revision |
---|
496 | */ |
---|
497 | public String getRevision() { |
---|
498 | return RevisionUtils.extract("$Revision: 5928 $"); |
---|
499 | } |
---|
500 | } |
---|
501 | |
---|
502 | /** |
---|
503 | * Returns default capabilities of the classifier. |
---|
504 | * |
---|
505 | * @return the capabilities of this classifier |
---|
506 | */ |
---|
507 | public Capabilities getCapabilities() { |
---|
508 | Capabilities result = super.getCapabilities(); |
---|
509 | result.disableAll(); |
---|
510 | |
---|
511 | // attributes |
---|
512 | result.enable(Capability.NOMINAL_ATTRIBUTES); |
---|
513 | result.enable(Capability.RELATIONAL_ATTRIBUTES); |
---|
514 | result.enable(Capability.MISSING_VALUES); |
---|
515 | |
---|
516 | // class |
---|
517 | result.enable(Capability.BINARY_CLASS); |
---|
518 | result.enable(Capability.MISSING_CLASS_VALUES); |
---|
519 | |
---|
520 | // other |
---|
521 | result.enable(Capability.ONLY_MULTIINSTANCE); |
---|
522 | |
---|
523 | return result; |
---|
524 | } |
---|
525 | |
---|
526 | /** |
---|
527 | * Returns the capabilities of this multi-instance classifier for the |
---|
528 | * relational data. |
---|
529 | * |
---|
530 | * @return the capabilities of this object |
---|
531 | * @see Capabilities |
---|
532 | */ |
---|
533 | public Capabilities getMultiInstanceCapabilities() { |
---|
534 | Capabilities result = super.getCapabilities(); |
---|
535 | result.disableAll(); |
---|
536 | |
---|
537 | // attributes |
---|
538 | result.enable(Capability.NOMINAL_ATTRIBUTES); |
---|
539 | result.enable(Capability.NUMERIC_ATTRIBUTES); |
---|
540 | result.enable(Capability.DATE_ATTRIBUTES); |
---|
541 | result.enable(Capability.MISSING_VALUES); |
---|
542 | |
---|
543 | // class |
---|
544 | result.disableAllClasses(); |
---|
545 | result.enable(Capability.NO_CLASS); |
---|
546 | |
---|
547 | return result; |
---|
548 | } |
---|
549 | |
---|
550 | /** |
---|
551 | * Builds the classifier |
---|
552 | * |
---|
553 | * @param train the training data to be used for generating the |
---|
554 | * boosted classifier. |
---|
555 | * @throws Exception if the classifier could not be built successfully |
---|
556 | */ |
---|
557 | public void buildClassifier(Instances train) throws Exception { |
---|
558 | // can classifier handle the data? |
---|
559 | getCapabilities().testWithFail(train); |
---|
560 | |
---|
561 | // remove instances with missing class |
---|
562 | train = new Instances(train); |
---|
563 | train.deleteWithMissingClass(); |
---|
564 | |
---|
565 | m_NumClasses = train.numClasses(); |
---|
566 | |
---|
567 | int nR = train.attribute(1).relation().numAttributes(); |
---|
568 | int nC = train.numInstances(); |
---|
569 | |
---|
570 | m_Data = new double [nC][nR][]; // Data values |
---|
571 | m_Classes = new int [nC]; // Class values |
---|
572 | m_Attributes = train.attribute(1).relation(); |
---|
573 | |
---|
574 | xMean = new double [nR]; // Mean of mean |
---|
575 | xSD = new double [nR]; // Mode of stddev |
---|
576 | |
---|
577 | double sY1=0, sY0=0, totIns=0; // Number of classes |
---|
578 | int[] missingbags = new int[nR]; |
---|
579 | |
---|
580 | if (m_Debug) { |
---|
581 | System.out.println("Extracting data..."); |
---|
582 | } |
---|
583 | |
---|
584 | for(int h=0; h<m_Data.length; h++){ |
---|
585 | Instance current = train.instance(h); |
---|
586 | m_Classes[h] = (int)current.classValue(); // Class value starts from 0 |
---|
587 | Instances currInsts = current.relationalValue(1); |
---|
588 | int nI = currInsts.numInstances(); |
---|
589 | totIns += (double)nI; |
---|
590 | |
---|
591 | for (int i = 0; i < nR; i++) { |
---|
592 | // initialize m_data[][][] |
---|
593 | m_Data[h][i] = new double[nI]; |
---|
594 | double avg=0, std=0, num=0; |
---|
595 | for (int k=0; k<nI; k++){ |
---|
596 | if(!currInsts.instance(k).isMissing(i)){ |
---|
597 | m_Data[h][i][k] = currInsts.instance(k).value(i); |
---|
598 | avg += m_Data[h][i][k]; |
---|
599 | std += m_Data[h][i][k]*m_Data[h][i][k]; |
---|
600 | num++; |
---|
601 | } |
---|
602 | else |
---|
603 | m_Data[h][i][k] = Double.NaN; |
---|
604 | } |
---|
605 | |
---|
606 | if(num > 0){ |
---|
607 | xMean[i] += avg/num; |
---|
608 | xSD[i] += std/num; |
---|
609 | } |
---|
610 | else |
---|
611 | missingbags[i]++; |
---|
612 | } |
---|
613 | |
---|
614 | // Class count |
---|
615 | if (m_Classes[h] == 1) |
---|
616 | sY1++; |
---|
617 | else |
---|
618 | sY0++; |
---|
619 | } |
---|
620 | |
---|
621 | for (int j = 0; j < nR; j++) { |
---|
622 | xMean[j] = xMean[j]/(double)(nC-missingbags[j]); |
---|
623 | xSD[j] = Math.sqrt(Math.abs(xSD[j]/((double)(nC-missingbags[j])-1.0) |
---|
624 | -xMean[j]*xMean[j]*(double)(nC-missingbags[j])/ |
---|
625 | ((double)(nC-missingbags[j])-1.0))); |
---|
626 | } |
---|
627 | |
---|
628 | if (m_Debug) { |
---|
629 | // Output stats about input data |
---|
630 | System.out.println("Descriptives..."); |
---|
631 | System.out.println(sY0 + " bags have class 0 and " + |
---|
632 | sY1 + " bags have class 1"); |
---|
633 | System.out.println("\n Variable Avg SD "); |
---|
634 | for (int j = 0; j < nR; j++) |
---|
635 | System.out.println(Utils.doubleToString(j,8,4) |
---|
636 | + Utils.doubleToString(xMean[j], 10, 4) |
---|
637 | + Utils.doubleToString(xSD[j], 10,4)); |
---|
638 | } |
---|
639 | |
---|
640 | // Normalise input data and remove ignored attributes |
---|
641 | for (int i = 0; i < nC; i++) { |
---|
642 | for (int j = 0; j < nR; j++) { |
---|
643 | for(int k=0; k < m_Data[i][j].length; k++){ |
---|
644 | if(xSD[j] != 0){ |
---|
645 | if(!Double.isNaN(m_Data[i][j][k])) |
---|
646 | m_Data[i][j][k] = (m_Data[i][j][k] - xMean[j]) / xSD[j]; |
---|
647 | else |
---|
648 | m_Data[i][j][k] = 0; |
---|
649 | } |
---|
650 | } |
---|
651 | } |
---|
652 | } |
---|
653 | |
---|
654 | if (m_Debug) { |
---|
655 | System.out.println("\nIteration History..." ); |
---|
656 | } |
---|
657 | |
---|
658 | double x[] = new double[nR + 1]; |
---|
659 | x[0] = Math.log((sY1+1.0) / (sY0+1.0)); |
---|
660 | double[][] b = new double[2][x.length]; |
---|
661 | b[0][0] = Double.NaN; |
---|
662 | b[1][0] = Double.NaN; |
---|
663 | for (int q=1; q < x.length;q++){ |
---|
664 | x[q] = 0.0; |
---|
665 | b[0][q] = Double.NaN; |
---|
666 | b[1][q] = Double.NaN; |
---|
667 | } |
---|
668 | |
---|
669 | OptEng opt = new OptEng(m_AlgorithmType); |
---|
670 | opt.setDebug(m_Debug); |
---|
671 | m_Par = opt.findArgmin(x, b); |
---|
672 | while(m_Par==null){ |
---|
673 | m_Par = opt.getVarbValues(); |
---|
674 | if (m_Debug) |
---|
675 | System.out.println("200 iterations finished, not enough!"); |
---|
676 | m_Par = opt.findArgmin(m_Par, b); |
---|
677 | } |
---|
678 | if (m_Debug) |
---|
679 | System.out.println(" -------------<Converged>--------------"); |
---|
680 | |
---|
681 | // feature selection use |
---|
682 | if (m_AlgorithmType == ALGORITHMTYPE_ARITHMETIC) { |
---|
683 | double[] fs = new double[nR]; |
---|
684 | for(int k=1; k<nR+1; k++) |
---|
685 | fs[k-1] = Math.abs(m_Par[k]); |
---|
686 | int[] idx = Utils.sort(fs); |
---|
687 | double max = fs[idx[idx.length-1]]; |
---|
688 | for(int k=idx.length-1; k>=0; k--) |
---|
689 | System.out.println(m_Attributes.attribute(idx[k]).name()+"\t"+(fs[idx[k]]*100/max)); |
---|
690 | } |
---|
691 | |
---|
692 | // Convert coefficients back to non-normalized attribute units |
---|
693 | for(int j = 1; j < nR+1; j++) { |
---|
694 | if (xSD[j-1] != 0) { |
---|
695 | m_Par[j] /= xSD[j-1]; |
---|
696 | m_Par[0] -= m_Par[j] * xMean[j-1]; |
---|
697 | } |
---|
698 | } |
---|
699 | } |
---|
700 | |
---|
701 | /** |
---|
702 | * Computes the distribution for a given exemplar |
---|
703 | * |
---|
704 | * @param exmp the exemplar for which distribution is computed |
---|
705 | * @return the distribution |
---|
706 | * @throws Exception if the distribution can't be computed successfully |
---|
707 | */ |
---|
708 | public double[] distributionForInstance(Instance exmp) |
---|
709 | throws Exception { |
---|
710 | |
---|
711 | // Extract the data |
---|
712 | Instances ins = exmp.relationalValue(1); |
---|
713 | int nI = ins.numInstances(), nA = ins.numAttributes(); |
---|
714 | double[][] dat = new double [nI][nA+1]; |
---|
715 | for(int j=0; j<nI; j++){ |
---|
716 | dat[j][0]=1.0; |
---|
717 | int idx=1; |
---|
718 | for(int k=0; k<nA; k++){ |
---|
719 | if(!ins.instance(j).isMissing(k)) |
---|
720 | dat[j][idx] = ins.instance(j).value(k); |
---|
721 | else |
---|
722 | dat[j][idx] = xMean[idx-1]; |
---|
723 | idx++; |
---|
724 | } |
---|
725 | } |
---|
726 | |
---|
727 | // Compute the probability of the bag |
---|
728 | double [] distribution = new double[2]; |
---|
729 | switch (m_AlgorithmType) { |
---|
730 | case ALGORITHMTYPE_DEFAULT: |
---|
731 | distribution[0]=0.0; // Log-Prob. for class 0 |
---|
732 | |
---|
733 | for(int i=0; i<nI; i++){ |
---|
734 | double exp = 0.0; |
---|
735 | for(int r=0; r<m_Par.length; r++) |
---|
736 | exp += m_Par[r]*dat[i][r]; |
---|
737 | exp = Math.exp(exp); |
---|
738 | |
---|
739 | // Prob. updated for one instance |
---|
740 | distribution[0] -= Math.log(1.0+exp); |
---|
741 | } |
---|
742 | |
---|
743 | // Prob. for class 0 |
---|
744 | distribution[0] = Math.exp(distribution[0]); |
---|
745 | // Prob. for class 1 |
---|
746 | distribution[1] = 1.0 - distribution[0]; |
---|
747 | break; |
---|
748 | |
---|
749 | case ALGORITHMTYPE_ARITHMETIC: |
---|
750 | distribution[0]=0.0; // Prob. for class 0 |
---|
751 | |
---|
752 | for(int i=0; i<nI; i++){ |
---|
753 | double exp = 0.0; |
---|
754 | for(int r=0; r<m_Par.length; r++) |
---|
755 | exp += m_Par[r]*dat[i][r]; |
---|
756 | exp = Math.exp(exp); |
---|
757 | |
---|
758 | // Prob. updated for one instance |
---|
759 | distribution[0] += 1.0/(1.0+exp); |
---|
760 | } |
---|
761 | |
---|
762 | // Prob. for class 0 |
---|
763 | distribution[0] /= (double)nI; |
---|
764 | // Prob. for class 1 |
---|
765 | distribution[1] = 1.0 - distribution[0]; |
---|
766 | break; |
---|
767 | |
---|
768 | case ALGORITHMTYPE_GEOMETRIC: |
---|
769 | for(int i=0; i<nI; i++){ |
---|
770 | double exp = 0.0; |
---|
771 | for(int r=0; r<m_Par.length; r++) |
---|
772 | exp += m_Par[r]*dat[i][r]; |
---|
773 | distribution[1] += exp/(double)nI; |
---|
774 | } |
---|
775 | |
---|
776 | // Prob. for class 1 |
---|
777 | distribution[1] = 1.0/(1.0+Math.exp(-distribution[1])); |
---|
778 | // Prob. for class 0 |
---|
779 | distribution[0] = 1-distribution[1]; |
---|
780 | break; |
---|
781 | } |
---|
782 | |
---|
783 | return distribution; |
---|
784 | } |
---|
785 | |
---|
786 | /** |
---|
787 | * Gets a string describing the classifier. |
---|
788 | * |
---|
789 | * @return a string describing the classifer built. |
---|
790 | */ |
---|
791 | public String toString() { |
---|
792 | |
---|
793 | String result = "Modified Logistic Regression"; |
---|
794 | if (m_Par == null) { |
---|
795 | return result + ": No model built yet."; |
---|
796 | } |
---|
797 | |
---|
798 | result += "\nMean type: " + getAlgorithmType().getSelectedTag().getReadable() + "\n"; |
---|
799 | result += "\nCoefficients...\n" |
---|
800 | + "Variable Coeff.\n"; |
---|
801 | for (int j = 1, idx=0; j < m_Par.length; j++, idx++) { |
---|
802 | result += m_Attributes.attribute(idx).name(); |
---|
803 | result += " "+Utils.doubleToString(m_Par[j], 12, 4); |
---|
804 | result += "\n"; |
---|
805 | } |
---|
806 | |
---|
807 | result += "Intercept:"; |
---|
808 | result += " "+Utils.doubleToString(m_Par[0], 10, 4); |
---|
809 | result += "\n"; |
---|
810 | |
---|
811 | result += "\nOdds Ratios...\n" |
---|
812 | + "Variable O.R.\n"; |
---|
813 | for (int j = 1, idx=0; j < m_Par.length; j++, idx++) { |
---|
814 | result += " " + m_Attributes.attribute(idx).name(); |
---|
815 | double ORc = Math.exp(m_Par[j]); |
---|
816 | result += " " + ((ORc > 1e10) ? "" + ORc : Utils.doubleToString(ORc, 12, 4)); |
---|
817 | } |
---|
818 | result += "\n"; |
---|
819 | return result; |
---|
820 | } |
---|
821 | |
---|
822 | /** |
---|
823 | * Returns the revision string. |
---|
824 | * |
---|
825 | * @return the revision |
---|
826 | */ |
---|
827 | public String getRevision() { |
---|
828 | return RevisionUtils.extract("$Revision: 5928 $"); |
---|
829 | } |
---|
830 | |
---|
831 | /** |
---|
832 | * Main method for testing this class. |
---|
833 | * |
---|
834 | * @param argv should contain the command line arguments to the |
---|
835 | * scheme (see Evaluation) |
---|
836 | */ |
---|
837 | public static void main(String[] argv) { |
---|
838 | runClassifier(new MILR(), argv); |
---|
839 | } |
---|
840 | } |
---|