Debian Codegraph @

inst/NEWS

raw
1 Wishlist (formerly TODO):
2
3 * Implement the new scheme of handling classwt in classification.
4
5 * Use more compact storage of proximity matrix.
6
7 ========================================================================
8 Changes in 4.7-1:
9
10 * The long-standing wishlist item of case weights (used only in sampling)
11 is now available (thanks to urging from Robert Tishirani and implementation
12 by Ethan Liaw).
13
14 Changes in 4.6-14:
15
16 * Native routines are now reigstered.
17
18 Changes in 4.6-13:
19
20 * The tie-breaking mechanisms in various places were not implemented correctly.
21 (Thanks to Adele Cutler and Susan Durham for reporting!)
22
23 * In unsupervised mode, the synthetic data were generated without updating the
24 index matrix that trees actually operate on.
25
26 * In rfcv(), the importance measures were taken from the first column of the
27 $importance component of the randomForest object. Now corrected to use
28 importance(..., type=1).
29
30 * Maximum tree size was computed incorrectly, leading to possibly smaller than
31 maxinum possible sized trees grown.
32
33 * In classification code, splits on categorical variables with >=32 levels were
34 not encoded correctly.
35
36 * Memory usage increase due to .C(..., DUP=FALSE) no longer being supported.
37 Work is underway to transition to .Call() to reduce memory usage.
38
39 Changes in 4.6-12:
40
41 * The "inbag" component of the randomForest object now records the number of
42 times an observation is included in the bootstrap sample, instead of just
43 indicating whether it is included or not.
44
45 * The "ncat" component of the "forest" component of the randomForest object
46 now always has names that are the column names of x.
47
48 * The "inbag" component is combined properly now in the combine() function.
49
50 * Housekeeping: Cleaned up the DESCRIPTION, NAMESPACE, and URLs in help files.
51
52 Changes in 4.6-10:
53
54 * Fixed bugs introduced in 4.6-8 and 4.6-9. Levels for ordered factors
55 in predictors are no longer checked for consistency between training
56 and test data in predict.randomForest(). In 4.6-9 the response is mean
57 centered before running the algorithm (for regression), but the mean was
58 not added back in the $y component of the return object from
59 randomForest().
60
61 Changes in 4.6-9:
62
63 * Categorical predictors are allowed to have up to 53 categories now.
64
65 Changes in 4.6-8:
66
67 * There was a bug in the handling of ordered factors in randomForest.formula().
68 (Reported by Barry Rawlingson.)
69
70 * Changed randomForest.formula() and predict.randomForest() so that
71 predict() without newdata when na.action=na.exclude pads the output
72 with NAs to match the rows in the original data (suggested by Barry
73 Rawlingson.).
74
75 * There was a long standing bug in randomForest(..., proximity=TRUE, oob.prox=FALSE)
76 in regression mode. (Thanks to Stefano Romano for the report.)
77
78 Change in 4.6-6:
79
80 * In randomForest(), proximity matrix is not computed correctly if oob.prox=TRUE.
81 (Thanks to Adele Cutler for the report!)
82
83 * In rfcv(), changed the sample-splitting for regression to handle cases
84 when there are many ties in the data.
85
86 * The calculation of importanceSD for classification was missing a divisor.
87 (Thanks to Abhishek Jaiantilal for the report.)
88
89 ? The simulated data in unsupervised randomForest() were not in the right range.
90 (Thanks to Abhishek Jaiantilal for the report.)
91
92 * In rfImpute(), padded 1e-8 to the division by sum of proximities to avoid
93 division by 0.
94
95 * Some R code clean up to keep R CMD check happy (complete argument names).
96
97 Change in 4.6-6:
98
99 * Fixed yet another bug in checking for missing classes in the in-bag sample.
100
101 Change in 4.6-5:
102
103 * Minor bug fix for drawing samples in classification that was introduced
104 in 4.6-4. (Thanks to Joran Elias for reporting.)
105
106 Changes in 4.6-4:
107
108 * Changed the error condition added in 4.6-3 to the case when there are fewer
109 than two classes present in the in-bag sample.
110
111 Changes in 4.6-3:
112
113 * Fixed bugs in the tie-breaking code in various places. (Thanks to
114 Abhishek Jaiantilal and Nathan Longbotham for the report and Abhishek
115 for the fix.)
116
117 * Throw error if some class has no data after 10 sampling attempts in
118 classification. (Thanks to Abhishek Jaiantilal for the report.)
119
120 Changes in 4.6-2:
121
122 * Part of the enhancement in 4.5-37 causes segfault in R-2.12.x. That part
123 has been reverted to the older code for the time being while the problem
124 is investigated further.
125
126 Changes in 4.6-1:
127
128 * The package now includes the rfcv() function for feature selection. See
129 the reference in the help page for details.
130
131 * predict.randomForest() was not retaining names of observations in some
132 cases.
133
134 Changes in 4.5-37:
135
136 * Many repeated calls to predict.randomForest() should run faster, thanks
137 to Philip Pham and Rory Martin who pointed out some unnecessary overhead.
138
139 Changes in 4.5-36:
140
141 * outlier() now works when the input matrix to randomForest() has no
142 row names. (Reported by Rau Carrio Gaspar.)
143
144 * na.roughfix() is now much faster on some data (fixed based on idea from
145 Hadley Wickham; problem reported by Mike Williamson.)
146
147 * Corrected typos in the description of categorical splits in ?getTree.
148
149 Changes in 4.5-35:
150
151 * Fixed an error in partialPlot.randomForest(). Now the partial
152 dependence plots for classification data should be more sensible.
153 (Thanks to Adele Cutler for the bug report and patch.)
154
155 * Re-worded part of the help pages regarding variable importance
156 calculation.
157
158 Changes in 4.5-34:
159
160 * Fixed infinite loop when randomForest() is called with non-null
161 maxnodes.
162
163 * Fixed a bug in margin.default() that gave nonsensical results.
164
165 Changes in 4.5-33:
166
167 * Fixed a _long standing_ bug (existed since the original Fortran) in
168 randomForest(): If importance=TRUE and proximity=TRUE, the proximity
169 matrix returned is incorrect. Those computed with importance=FALSE, or
170 with predict.randomForest(..., proximity=TRUE) are correct.
171
172 Changes in 4.5-32:
173
174 * Fixed a bug in predict.randomForest(..., predict.all=TRUE) introduced in
175 4.5-31. Added examples in ?predict.randomForest for the options.
176
177 Changes in 4.5-31:
178
179 * Added a new option `maxnodes' in randomForest() that limits the size of
180 trees.
181
182 * margin() is now generic with a method for randomForest objects.
183
184 * Fixed the help page for getTree() about how data are split on numeric
185 variables (`<=' instead of `<').
186
187 * Fixed predict.randomForest() so that if the randomForest object is of
188 type "regression" and built from the formula interface and newdata
189 contains NAs, NAs are returned in the corresponding positions (instead
190 of being dropped altogether).
191
192 Change in 4.5-30:
193
194 * In regression, cases that had not been out-of-bag now gets NA as prediction
195 (as in classification).
196
197 Change in 4.5-29:
198
199 * Fixed a couple of benign errors in help pages spotted by the new Rd parser.
200
201 Change in 4.5-28:
202
203 * randomForest() would segfault if there are 32-level factors among the
204 predictor variables.
205
206 Changes in 4.5-27:
207
208 * Fixed handling of ordered factors in predict.randomForest().
209
210 Changes in 4.5-26:
211
212 * Fix formula parsing so use of functions in formula won't trigger errors.
213
214 * predict.randomForest() did not work when given a matrix without column
215 names as the newdata.
216
217 Changes in 4.5-25:
218
219 * In regression, the out-of-bag estimate of MSE and R-squared for the
220 first few trees (for which not all observations have been OOB yet),
221 were computed wrong, leading to gross over-estimates of MSEs for the
222 first few trees. This did not affect the final MSE and R-squared
223 for the whole forest; i.e., the first few elements of the `mse'
224 component (and thus the corresponding elements in `rsq') in the
225 randomForest object are wrong, but others are correct.
226
227 (Thanks to Ulrike Gromping for pointing out this problem, as well
228 as the one fixed in 4.5-24.)
229
230 Changes in 4.5-24:
231
232 * randomForest.formula() did not exclude terms preceded by -.
233
234 Changes in 4.5-23:
235
236 * Fixed tuneRF() to work with R version > 2.6.1.
237
238 * Make predict.randomForest() more backward compatible with randomForest
239 objects created from versions older than 4.5-21.
240
241 Changes in 4.5-22:
242
243 * Allow unsupervised randomForest not to produce a proximity matrix (by
244 specifying proximity=FALSE), suggested by Nick Crookston.
245
246 Changes in 4.5-21:
247
248 * The added check for factor level consistency in predictor variables in
249 4.5-20 was not working for predictors given in a matrix (reported by
250 Ramon Diaz-Uriate).
251
252 Changes in 4.5-20:
253
254 * Fixed a memory bug in the C code when the test set is given and
255 proximity is requested in regression. (Reported by Clayton Springer.)
256
257 * Fixed the one-pass random tie-breaking algorithm in various places.
258
259 * Added code to check consistency of levels for factors in the predictors,
260 as well as allowing missing levels of factors and extraneous variables in
261 predict(..., newdata). (Thanks to Nick Crookston for suggesting a patch.)
262
263 Changes in 4.5-19:
264
265 * In classification, if sampsize is small and sampling is not stratified,
266 the actual sample might be larger than specified in some trees. Now fixed.
267
268 * Fixed combine() to work on regression randomForest objects and for
269 cases when ntree is small.
270
271 * randomForest.default() for regression was unnecessarily creating
272 a matrix of 0s for localImportance when importance=TRUE but
273 localImp=FALSE. (Thanks to Robert McGehee for reporting these
274 bugs.)
275
276 * predict.randomForest(..., nodes=TRUE) now works for regression.
277
278 Changes in 4.5-18:
279
280 * Added S-PLUS 8 compatibility.
281
282 Changes in 4.5-17:
283
284 * Added `w' (for weights) to partialPlot.randomForest().
285
286 Changes in 4.5-16:
287
288 * Fixed some typos in the documentation source files (e.g., \note vs. \notes,
289 etc).
290
291 Changes in 4.5-15:
292
293 * Fixed error message call in predict.randomForest().
294
295 Changes in 4.5-14:
296
297 * varImpPlot() was ignoring the `type' argument.
298
299 * "<" was used instead of ".lt." in Fortran code, which is not F77-compliant.
300
301 Changes in 4.5-13:
302
303 * Fixed a bug in randomForest() when biasCorr=TRUE for regression.
304
305 * Fixed bug in predict.randomForest() when newdata is a matrix with no rownames.
306 Changes in 4.5-12:
307
308 * Added the `strata' argument to randomForest, which, in conjunction with
309 `sampsize', allow sampling (with or without replacement) according to a
310 strata variable (which can be something other than the class variable).
311 Currently only works in classification.
312
313 Changes in 4.5-11:
314
315 * Fixed partialPlot.randomForest() so that if x.var is a character, it's
316 taken as the name of the variable.
317
318 * Clean up code for importance() and varImpPlot() so that if the
319 randomForest object only contains one importance measure, varImpPlot()
320 will work as intended.
321
322 Changes in 4.5-10:
323
324 * Renamed the first argument of randomForest.formula() to `formula', to be
325 consistent with other formula interfaces.
326
327 Changes in 4.5-9:
328
329 * Fixed a bug with unsupervised randomForest(..., keep.forest=TRUE).
330
331 * Fixed a bug in regression that caused crash when proximiy=TRUE.
332
333 * Added `keep.inbag' argument to randomForest(), which, if set to TRUE,
334 cause randomForest() to return a matrix of indicators that indicate
335 which case is included in the bootstrap sample to grow the trees.
336
337 Changes in 4.5-8:
338
339 * Added some code in predict.randomForest() so it works with randomForest
340 objects created in older versions of the package.
341
342 * Fixed randomForest.default() so that getTree() works when the forest
343 contains only one tree.
344
345 * Added the argument `labelVar' (default FALSE) to getTree() for prettier
346 output.
347
348 Changes in 4.5-7:
349
350 * Fixed (another!) bug in splitting on categorical variables, especially
351 impacting data with binary (categorical) variables.
352
353 Changes in 4.5-6:
354
355 * Fixed a bug introduced in 4.5-2 that used the wrong default class weights.
356
357 Changes in 4.5-5:
358
359 * Fixed a couple of bugs in C/Fortran code for splitting on categorical
360 variables in classification trees, which lead to negative or Inf
361 decrease in the Gini index.
362
363 Changes in 4.5-4:
364
365 * Fixed a bug in regression when there are categorical predictors. (The
366 splits can be completely wrong!)
367
368 Changes in 4.5-3:
369
370 * Fixed predict.randomForest() so that it uses the class labels stored in
371 the randomForest object (for classification).
372
373 Changes in 4.5-2:
374
375 * New argument `cutoff' added to predict.randomForest(). The usage is
376 analogous to the same argument to randomForest().
377
378 * Added `palette' and `pch' arguments to MDSplot() to allow more user control.
379
380 * In randomForest(), allow the forest to be returned in `unsupervised' mode.
381
382 * Fixed some inaccuracies in help pages.
383
384 * Fixed the way version number of the package is found at start-up.
385
386 Changes in 4.5-1:
387
388 * In classification, split on a categorical predictor with more than 10
389 categories is made more efficient: For two-class problems, the
390 heuristic presented in Section 4.2.2 of the CART book is used.
391 Otherwise 512 randomly sampled (not necessarily unique) splits are
392 tested, instead of all possible splits.
393
394 * New function classCenter() has been added. It takes a proximity matrix and
395 a vector of class labels and compute one prototype per class.
396
397 * Added the `Automobile' data from UCI Machine Learning Repository.
398
399 * Fixed partialPlot() for categorical predictors (wrong barplot was produced).
400
401 * Some re-organization and clean-up of internal C/Fortran code is on-going.
402
403 Changes in 4.4-3:
404
405 * Added the nPerm argument to randomForest(), which controls the number
406 of times the out-of-bag part of each variable is permuted, per tree,
407 for computing variable importance. (Currently only implemented for
408 regression.)
409
410 * When computing the out-of-bag MSE for each tree for assessing variable
411 importance in regression, the total number of cases was wrongly used as
412 the divisor.
413
414 * Fixed the default and formula methods of randomForest(), so that the
415 `call' component of the returned object calls the generic.
416
417 * The `% increase in MSE' measure of variable importance in regression was
418 not being computed correctly (should divide sum of squares by number of
419 out-of-bag samples rather than total number of samples, for each tree).
420
421 * Fixed a bug in na.roughfix.default() that gave warning with matrix input.
422
423 Changes in 4.4-2:
424
425 * Fixed two memory leaks in the regression code (introduced in 4.3-1).
426
427 * Fixed a bug that sometimes caused crash in regression when nodesize
428 is set to something larger than default (5).
429
430 * Changed the tree structure in regression slightly: "treemap" is replaced
431 by "leftDaughter" and "rightDaughter".
432
433 Changes in 4.4-1:
434
435 * Made slight change in regression code so that it won't split `pure'
436 nodes. Also fixed the `increase in node purity' importance measure
437 in regression.
438
439 * The outscale option in randomForest() is removed. Use the outlier()
440 function instead. The default outlier() method can be used with other
441 proximity/dissimilarity measures.
442
443 * More Fortran subroutines migrated to C.
444
445 Changes in 4.3-3:
446
447 * Fixed randomForest.formula() so that update() will work.
448
449 * Fixed up problem in importance(), which was broken in a couple of ways.
450
451 Changes in 4.3-2:
452
453 * Fixed a bug that caused crashes in classification if test set data
454 are supplied.
455
456 Changes in 4.3-1:
457
458 * Fixed bugs in sampling cases and variables without replacement.
459
460 * Added the rfNews() function to display the NEWS file. Advertised in
461 the start up banner.
462
463 * (Not user-visible.) Translated regression tree building code from
464 Fortran to C. One perhaps noticeable change is less memory usage.
465
466 Changes in 4.3-0:
467
468 * Thanks to Adele Cutler, there's now casewise variable importance
469 measures in classification. Similar feature is also added for
470 regression. Use the new localImp option in randomForest().
471
472 * The `importance' component of randomForest object has been changed:
473 The permutation-based measures are not divided by their `standard
474 errors'. Instead, the `standard errors' are stored in the
475 `importanceSD' component. One should use the importance() extractor
476 function rather than something like rf.obj$importance for extracting
477 the importance measures.
478
479 * The importance() extractor function has been updated: If the
480 permutation-based measures are available, calling importance()
481 with only a randomForest object returns the matrix of variable
482 importance measures. There is the `scale' argument, which defaults
483 to TRUE.
484
485 * In predict.randomForest, there is a new argument `nodes' (default to
486 FALSE). For classification, if nodes=TRUE, the returned object has an
487 attribute `nodes', which is an n by ntree matrix of terminal node
488 indicators. This is ignored for regression.
489
490 Changes in 4.2-1:
491
492 * There is now a package name space. Only generics are exported.
493
494 * Some function names have been changed:
495 partial.plot -> partialPlot
496 var.imp.plot -> varImpPlot
497 var.used -> varUsed
498
499 * There is a new option `replace' in randomForest() (default to TRUE)
500 indicating whether the sampling of cases is with or without
501 replacement.
502
503 * In randomForest(), the `sampsize' option now works for both
504 classification and regression, and indicate the number of cases to be
505 drawn to grow each tree. For classification, if sampsize is a vector of
506 length the number of classes, then sampling is stratified by class.
507
508 * With the formula interface for randomForest(), the default na.action,
509 na.fail, is effective. I.e., an error is given if there are NAs present
510 in the data. If na.omit is desired, it must be given explicitly.
511
512 * For classification, the err.rate component of the randomForest object
513 (and the corresponding one for test set) now is a ntree by (nclass + 1)
514 matrix, the first column of which contains the overall error rate, and
515 the remaining columns the class error rates. The running output now
516 also prints class error rates. The plot method for randomForest will
517 plot the class error rates as well.
518
519 * The predict() method now checks whether the variable names in newdata
520 match those from the training data (if the randomForest object is not
521 created from the formula interface).
522
523 * partialPlot() and varImpPlot() now have optional arguments xlab, ylab
524 and main for more flexible labelling. Also, if a factor is given as
525 the variable, a real bar plot is produced.
526
527 * partialPlot() will now remove rows with NAs from the data frame given.
528
529 * For regression, if proximity=FALSE, an n by n array of integers is
530 erroneously allocated but not used (it's only used for proximity
531 calculation, so not needed otherwise).
532
533 * Updated combine() to conform to the new randomForest object.
534
535 * na.roughfix() was not working correctly for matrices, which in turns
536 causes problem in rfImpute().
537
538
539 Changes in 4.1-0:
540
541 * In randomForest(), if sampsize is given, the sampling is now done
542 without replacement, in addition to stratified by class. Therefore
543 sampsize can not be larger than the class frequencies.
544
545 * In classification randomForest, checks are added to avoid trees with
546 only the root node.
547
548 * Fixed a bug in the Fortran code for classification that caused segfault
549 on some system when encountering a tree with only root node.
550
551 * The help page for predict.randomForest() now states the fact that when
552 newdata is not specified, the OOB predictions from the randomForest
553 object is returned.
554
555 * plot.randomForest() and print.randomForest() were not checking for
556 existence of performance (err.rate or mse) on test data correctly.