1
|
|
|
<?php |
2
|
|
|
|
3
|
|
|
declare(strict_types=1); |
4
|
|
|
|
5
|
|
|
namespace Phpml\Classification; |
6
|
|
|
|
7
|
|
|
use Phpml\Helper\Predictable; |
8
|
|
|
use Phpml\Helper\Trainable; |
9
|
|
|
use Phpml\Math\Statistic\Mean; |
10
|
|
|
use Phpml\Classification\DecisionTree\DecisionTreeLeaf; |
11
|
|
|
|
12
|
|
|
class DecisionTree implements Classifier |
13
|
|
|
{ |
14
|
|
|
use Trainable, Predictable; |
15
|
|
|
|
16
|
|
|
const CONTINUOS = 1; |
17
|
|
|
const NOMINAL = 2; |
18
|
|
|
|
19
|
|
|
/** |
20
|
|
|
* @var array |
21
|
|
|
*/ |
22
|
|
|
private $samples = []; |
23
|
|
|
|
24
|
|
|
/** |
25
|
|
|
* @var array |
26
|
|
|
*/ |
27
|
|
|
private $columnTypes; |
28
|
|
|
|
29
|
|
|
/** |
30
|
|
|
* @var array |
31
|
|
|
*/ |
32
|
|
|
private $labels = []; |
33
|
|
|
|
34
|
|
|
/** |
35
|
|
|
* @var int |
36
|
|
|
*/ |
37
|
|
|
private $featureCount = 0; |
38
|
|
|
|
39
|
|
|
/** |
40
|
|
|
* @var DecisionTreeLeaf |
41
|
|
|
*/ |
42
|
|
|
private $tree = null; |
43
|
|
|
|
44
|
|
|
/** |
45
|
|
|
* @var int |
46
|
|
|
*/ |
47
|
|
|
private $maxDepth; |
48
|
|
|
|
49
|
|
|
/** |
50
|
|
|
* @var int |
51
|
|
|
*/ |
52
|
|
|
public $actualDepth = 0; |
53
|
|
|
|
54
|
|
|
/** |
55
|
|
|
* @var int |
56
|
|
|
*/ |
57
|
|
|
private $numUsableFeatures = 0; |
58
|
|
|
|
59
|
|
|
/** |
60
|
|
|
* @var array |
61
|
|
|
*/ |
62
|
|
|
private $featureImportances = null; |
63
|
|
|
|
64
|
|
|
/** |
65
|
|
|
* |
66
|
|
|
* @var array |
67
|
|
|
*/ |
68
|
|
|
private $columnNames = null; |
69
|
|
|
|
70
|
|
|
/** |
71
|
|
|
* @param int $maxDepth |
72
|
|
|
*/ |
73
|
|
|
public function __construct($maxDepth = 10) |
74
|
|
|
{ |
75
|
|
|
$this->maxDepth = $maxDepth; |
76
|
|
|
} |
77
|
|
|
/** |
78
|
|
|
* @param array $samples |
79
|
|
|
* @param array $targets |
80
|
|
|
*/ |
81
|
|
|
public function train(array $samples, array $targets) |
82
|
|
|
{ |
83
|
|
|
$this->samples = array_merge($this->samples, $samples); |
84
|
|
|
$this->targets = array_merge($this->targets, $targets); |
85
|
|
|
|
86
|
|
|
$this->featureCount = count($this->samples[0]); |
87
|
|
|
$this->columnTypes = $this->getColumnTypes($this->samples); |
88
|
|
|
$this->labels = array_keys(array_count_values($this->targets)); |
89
|
|
|
$this->tree = $this->getSplitLeaf(range(0, count($this->samples) - 1)); |
90
|
|
|
|
91
|
|
|
// Each time the tree is trained, feature importances are reset so that |
92
|
|
|
// we will have to compute it again depending on the new data |
93
|
|
|
$this->featureImportances = null; |
|
|
|
|
94
|
|
|
|
95
|
|
|
// If column names are given or computed before, then there is no |
96
|
|
|
// need to init it and accidentally remove the previous given names |
97
|
|
|
if ($this->columnNames === null) { |
98
|
|
|
$this->columnNames = range(0, $this->featureCount - 1); |
99
|
|
|
} elseif (count($this->columnNames) > $this->featureCount) { |
100
|
|
|
$this->columnNames = array_slice($this->columnNames, 0, $this->featureCount); |
101
|
|
|
} elseif (count($this->columnNames) < $this->featureCount) { |
102
|
|
|
$this->columnNames = array_merge($this->columnNames, |
103
|
|
|
range(count($this->columnNames), $this->featureCount - 1)); |
104
|
|
|
} |
105
|
|
|
} |
106
|
|
|
|
107
|
|
|
protected function getColumnTypes(array $samples) |
108
|
|
|
{ |
109
|
|
|
$types = []; |
110
|
|
|
for ($i=0; $i<$this->featureCount; $i++) { |
111
|
|
|
$values = array_column($samples, $i); |
112
|
|
|
$isCategorical = $this->isCategoricalColumn($values); |
113
|
|
|
$types[] = $isCategorical ? self::NOMINAL : self::CONTINUOS; |
114
|
|
|
} |
115
|
|
|
return $types; |
116
|
|
|
} |
117
|
|
|
|
118
|
|
|
/** |
119
|
|
|
* @param null|array $records |
120
|
|
|
* @return DecisionTreeLeaf |
121
|
|
|
*/ |
122
|
|
|
protected function getSplitLeaf($records, $depth = 0) |
123
|
|
|
{ |
124
|
|
|
$split = $this->getBestSplit($records); |
|
|
|
|
125
|
|
|
$split->level = $depth; |
126
|
|
|
if ($this->actualDepth < $depth) { |
127
|
|
|
$this->actualDepth = $depth; |
128
|
|
|
} |
129
|
|
|
$leftRecords = []; |
130
|
|
|
$rightRecords= []; |
131
|
|
|
$remainingTargets = []; |
132
|
|
|
$prevRecord = null; |
133
|
|
|
$allSame = true; |
134
|
|
|
foreach ($records as $recordNo) { |
|
|
|
|
135
|
|
|
$record = $this->samples[$recordNo]; |
136
|
|
|
if ($prevRecord && $prevRecord != $record) { |
137
|
|
|
$allSame = false; |
138
|
|
|
} |
139
|
|
|
$prevRecord = $record; |
140
|
|
|
if ($split->evaluate($record)) { |
141
|
|
|
$leftRecords[] = $recordNo; |
142
|
|
|
} else { |
143
|
|
|
$rightRecords[]= $recordNo; |
144
|
|
|
} |
145
|
|
|
$target = $this->targets[$recordNo]; |
146
|
|
|
if (! in_array($target, $remainingTargets)) { |
147
|
|
|
$remainingTargets[] = $target; |
148
|
|
|
} |
149
|
|
|
} |
150
|
|
|
|
151
|
|
|
if (count($remainingTargets) == 1 || $allSame || $depth >= $this->maxDepth) { |
152
|
|
|
$split->isTerminal = 1; |
|
|
|
|
153
|
|
|
$classes = array_count_values($remainingTargets); |
154
|
|
|
arsort($classes); |
155
|
|
|
$split->classValue = key($classes); |
156
|
|
|
} else { |
157
|
|
|
if ($leftRecords) { |
|
|
|
|
158
|
|
|
$split->leftLeaf = $this->getSplitLeaf($leftRecords, $depth + 1); |
159
|
|
|
} |
160
|
|
|
if ($rightRecords) { |
|
|
|
|
161
|
|
|
$split->rightLeaf= $this->getSplitLeaf($rightRecords, $depth + 1); |
162
|
|
|
} |
163
|
|
|
} |
164
|
|
|
return $split; |
165
|
|
|
} |
166
|
|
|
|
167
|
|
|
/** |
168
|
|
|
* @param array $records |
169
|
|
|
* @return DecisionTreeLeaf[] |
170
|
|
|
*/ |
171
|
|
|
protected function getBestSplit($records) |
172
|
|
|
{ |
173
|
|
|
$targets = array_intersect_key($this->targets, array_flip($records)); |
174
|
|
|
$samples = array_intersect_key($this->samples, array_flip($records)); |
175
|
|
|
$samples = array_combine($records, $this->preprocess($samples)); |
176
|
|
|
$bestGiniVal = 1; |
177
|
|
|
$bestSplit = null; |
178
|
|
|
$features = $this->getSelectedFeatures(); |
179
|
|
|
foreach ($features as $i) { |
180
|
|
|
$colValues = []; |
181
|
|
|
foreach ($samples as $index => $row) { |
182
|
|
|
$colValues[$index] = $row[$i]; |
183
|
|
|
} |
184
|
|
|
$counts = array_count_values($colValues); |
185
|
|
|
arsort($counts); |
186
|
|
|
$baseValue = key($counts); |
187
|
|
|
$gini = $this->getGiniIndex($baseValue, $colValues, $targets); |
188
|
|
|
if ($bestSplit == null || $bestGiniVal > $gini) { |
189
|
|
|
$split = new DecisionTreeLeaf(); |
190
|
|
|
$split->value = $baseValue; |
191
|
|
|
$split->giniIndex = $gini; |
|
|
|
|
192
|
|
|
$split->columnIndex = $i; |
193
|
|
|
$split->isContinuous = $this->columnTypes[$i] == self::CONTINUOS; |
194
|
|
|
$split->records = $records; |
195
|
|
|
$bestSplit = $split; |
196
|
|
|
$bestGiniVal = $gini; |
197
|
|
|
} |
198
|
|
|
} |
199
|
|
|
return $bestSplit; |
200
|
|
|
} |
201
|
|
|
|
202
|
|
|
/** |
203
|
|
|
* @return array |
204
|
|
|
*/ |
205
|
|
|
protected function getSelectedFeatures() |
206
|
|
|
{ |
207
|
|
|
$allFeatures = range(0, $this->featureCount - 1); |
208
|
|
|
if ($this->numUsableFeatures == 0) { |
209
|
|
|
return $allFeatures; |
210
|
|
|
} |
211
|
|
|
|
212
|
|
|
$numFeatures = $this->numUsableFeatures; |
213
|
|
|
if ($numFeatures > $this->featureCount) { |
214
|
|
|
$numFeatures = $this->featureCount; |
215
|
|
|
} |
216
|
|
|
shuffle($allFeatures); |
217
|
|
|
$selectedFeatures = array_slice($allFeatures, 0, $numFeatures, false); |
218
|
|
|
sort($selectedFeatures); |
219
|
|
|
|
220
|
|
|
return $selectedFeatures; |
221
|
|
|
} |
222
|
|
|
|
223
|
|
|
/** |
224
|
|
|
* @param string $baseValue |
225
|
|
|
* @param array $colValues |
226
|
|
|
* @param array $targets |
227
|
|
|
*/ |
228
|
|
|
public function getGiniIndex($baseValue, $colValues, $targets) |
229
|
|
|
{ |
230
|
|
|
$countMatrix = []; |
231
|
|
|
foreach ($this->labels as $label) { |
232
|
|
|
$countMatrix[$label] = [0, 0]; |
233
|
|
|
} |
234
|
|
|
foreach ($colValues as $index => $value) { |
235
|
|
|
$label = $targets[$index]; |
236
|
|
|
$rowIndex = $value == $baseValue ? 0 : 1; |
237
|
|
|
$countMatrix[$label][$rowIndex]++; |
238
|
|
|
} |
239
|
|
|
$giniParts = [0, 0]; |
240
|
|
|
for ($i=0; $i<=1; $i++) { |
241
|
|
|
$part = 0; |
242
|
|
|
$sum = array_sum(array_column($countMatrix, $i)); |
243
|
|
|
if ($sum > 0) { |
244
|
|
|
foreach ($this->labels as $label) { |
245
|
|
|
$part += pow($countMatrix[$label][$i] / floatval($sum), 2); |
246
|
|
|
} |
247
|
|
|
} |
248
|
|
|
$giniParts[$i] = (1 - $part) * $sum; |
249
|
|
|
} |
250
|
|
|
return array_sum($giniParts) / count($colValues); |
251
|
|
|
} |
252
|
|
|
|
253
|
|
|
/** |
254
|
|
|
* @param array $samples |
255
|
|
|
* @return array |
256
|
|
|
*/ |
257
|
|
|
protected function preprocess(array $samples) |
258
|
|
|
{ |
259
|
|
|
// Detect and convert continuous data column values into |
260
|
|
|
// discrete values by using the median as a threshold value |
261
|
|
|
$columns = []; |
262
|
|
|
for ($i=0; $i<$this->featureCount; $i++) { |
263
|
|
|
$values = array_column($samples, $i); |
264
|
|
|
if ($this->columnTypes[$i] == self::CONTINUOS) { |
265
|
|
|
$median = Mean::median($values); |
266
|
|
|
foreach ($values as &$value) { |
267
|
|
|
if ($value <= $median) { |
268
|
|
|
$value = "<= $median"; |
269
|
|
|
} else { |
270
|
|
|
$value = "> $median"; |
271
|
|
|
} |
272
|
|
|
} |
273
|
|
|
} |
274
|
|
|
$columns[] = $values; |
275
|
|
|
} |
276
|
|
|
// Below method is a strange yet very simple & efficient method |
277
|
|
|
// to get the transpose of a 2D array |
278
|
|
|
return array_map(null, ...$columns); |
279
|
|
|
} |
280
|
|
|
|
281
|
|
|
/** |
282
|
|
|
* @param array $columnValues |
283
|
|
|
* @return bool |
284
|
|
|
*/ |
285
|
|
|
protected function isCategoricalColumn(array $columnValues) |
286
|
|
|
{ |
287
|
|
|
$count = count($columnValues); |
288
|
|
|
// There are two main indicators that *may* show whether a |
289
|
|
|
// column is composed of discrete set of values: |
290
|
|
|
// 1- Column may contain string values |
291
|
|
|
// 2- Number of unique values in the column is only a small fraction of |
292
|
|
|
// all values in that column (Lower than or equal to %20 of all values) |
293
|
|
|
$numericValues = array_filter($columnValues, 'is_numeric'); |
294
|
|
|
if (count($numericValues) != $count) { |
295
|
|
|
return true; |
296
|
|
|
} |
297
|
|
|
$distinctValues = array_count_values($columnValues); |
298
|
|
|
if (count($distinctValues) <= $count / 5) { |
299
|
|
|
return true; |
300
|
|
|
} |
301
|
|
|
return false; |
302
|
|
|
} |
303
|
|
|
|
304
|
|
|
/** |
305
|
|
|
* This method is used to set number of columns to be used |
306
|
|
|
* when deciding a split at an internal node of the tree. <br> |
307
|
|
|
* If the value is given 0, then all features are used (default behaviour), |
308
|
|
|
* otherwise the given value will be used as a maximum for number of columns |
309
|
|
|
* randomly selected for each split operation. |
310
|
|
|
* |
311
|
|
|
* @param int $numFeatures |
312
|
|
|
* @return $this |
313
|
|
|
* @throws Exception |
314
|
|
|
*/ |
315
|
|
|
public function setNumFeatures(int $numFeatures) |
316
|
|
|
{ |
317
|
|
|
if ($numFeatures < 0) { |
318
|
|
|
throw new \Exception("Selected column count should be greater or equal to zero"); |
319
|
|
|
} |
320
|
|
|
|
321
|
|
|
$this->numUsableFeatures = $numFeatures; |
322
|
|
|
|
323
|
|
|
return $this; |
324
|
|
|
} |
325
|
|
|
|
326
|
|
|
/** |
327
|
|
|
* A string array to represent columns. Useful when HTML output or |
328
|
|
|
* column importances are desired to be inspected. |
329
|
|
|
* |
330
|
|
|
* @param array $names |
331
|
|
|
* @return $this |
332
|
|
|
*/ |
333
|
|
|
public function setColumnNames(array $names) |
334
|
|
|
{ |
335
|
|
|
if ($this->featureCount != 0 && count($names) != $this->featureCount) { |
336
|
|
|
throw new \Exception("Length of the given array should be equal to feature count ($this->featureCount)"); |
337
|
|
|
} |
338
|
|
|
|
339
|
|
|
$this->columnNames = $names; |
340
|
|
|
|
341
|
|
|
return $this; |
342
|
|
|
} |
343
|
|
|
|
344
|
|
|
/** |
345
|
|
|
* @return string |
346
|
|
|
*/ |
347
|
|
|
public function getHtml() |
348
|
|
|
{ |
349
|
|
|
return $this->tree->getHTML($this->columnNames); |
350
|
|
|
} |
351
|
|
|
|
352
|
|
|
/** |
353
|
|
|
* This will return an array including an importance value for |
354
|
|
|
* each column in the given dataset. The importance values are |
355
|
|
|
* normalized and their total makes 1.<br/> |
356
|
|
|
* |
357
|
|
|
* @param array $labels |
|
|
|
|
358
|
|
|
* @return array |
359
|
|
|
*/ |
360
|
|
|
public function getFeatureImportances() |
361
|
|
|
{ |
362
|
|
|
if ($this->featureImportances !== null) { |
363
|
|
|
return $this->featureImportances; |
364
|
|
|
} |
365
|
|
|
|
366
|
|
|
$sampleCount = count($this->samples); |
367
|
|
|
$this->featureImportances = []; |
368
|
|
|
foreach ($this->columnNames as $column => $columnName) { |
369
|
|
|
$nodes = $this->getSplitNodesByColumn($column, $this->tree); |
370
|
|
|
|
371
|
|
|
$importance = 0; |
372
|
|
|
foreach ($nodes as $node) { |
373
|
|
|
$importance += $node->getNodeImpurityDecrease($sampleCount); |
374
|
|
|
} |
375
|
|
|
|
376
|
|
|
$this->featureImportances[$columnName] = $importance; |
377
|
|
|
} |
378
|
|
|
|
379
|
|
|
// Normalize & sort the importances |
380
|
|
|
$total = array_sum($this->featureImportances); |
381
|
|
|
if ($total > 0) { |
382
|
|
|
foreach ($this->featureImportances as &$importance) { |
383
|
|
|
$importance /= $total; |
384
|
|
|
} |
385
|
|
|
arsort($this->featureImportances); |
386
|
|
|
} |
387
|
|
|
|
388
|
|
|
return $this->featureImportances; |
389
|
|
|
} |
390
|
|
|
|
391
|
|
|
/** |
392
|
|
|
* Collects and returns an array of internal nodes that use the given |
393
|
|
|
* column as a split criteron |
394
|
|
|
* |
395
|
|
|
* @param int $column |
396
|
|
|
* @param DecisionTreeLeaf |
397
|
|
|
* @param array $collected |
|
|
|
|
398
|
|
|
* |
399
|
|
|
* @return array |
400
|
|
|
*/ |
401
|
|
|
protected function getSplitNodesByColumn($column, DecisionTreeLeaf $node) |
402
|
|
|
{ |
403
|
|
|
if (!$node || $node->isTerminal) { |
404
|
|
|
return []; |
405
|
|
|
} |
406
|
|
|
|
407
|
|
|
$nodes = []; |
408
|
|
|
if ($node->columnIndex == $column) { |
409
|
|
|
$nodes[] = $node; |
410
|
|
|
} |
411
|
|
|
|
412
|
|
|
$lNodes = []; |
413
|
|
|
$rNodes = []; |
414
|
|
|
if ($node->leftLeaf) { |
415
|
|
|
$lNodes = $this->getSplitNodesByColumn($column, $node->leftLeaf); |
416
|
|
|
} |
417
|
|
|
if ($node->rightLeaf) { |
418
|
|
|
$rNodes = $this->getSplitNodesByColumn($column, $node->rightLeaf); |
419
|
|
|
} |
420
|
|
|
$nodes = array_merge($nodes, $lNodes, $rNodes); |
421
|
|
|
|
422
|
|
|
return $nodes; |
423
|
|
|
} |
424
|
|
|
|
425
|
|
|
/** |
426
|
|
|
* @param array $sample |
427
|
|
|
* @return mixed |
428
|
|
|
*/ |
429
|
|
|
protected function predictSample(array $sample) |
430
|
|
|
{ |
431
|
|
|
$node = $this->tree; |
432
|
|
|
do { |
433
|
|
|
if ($node->isTerminal) { |
434
|
|
|
break; |
435
|
|
|
} |
436
|
|
|
if ($node->evaluate($sample)) { |
437
|
|
|
$node = $node->leftLeaf; |
438
|
|
|
} else { |
439
|
|
|
$node = $node->rightLeaf; |
440
|
|
|
} |
441
|
|
|
} while ($node); |
442
|
|
|
|
443
|
|
|
return $node ? $node->classValue : $this->labels[0]; |
444
|
|
|
} |
445
|
|
|
} |
446
|
|
|
|
Our type inference engine has found an assignment to a property that is incompatible with the declared type of that property.
Either this assignment is in error or the assigned type should be added to the documentation/type hint for that property..