1
|
|
|
<?php |
2
|
|
|
|
3
|
|
|
declare(strict_types=1); |
4
|
|
|
|
5
|
|
|
namespace Phpml\Clustering; |
6
|
|
|
|
7
|
|
|
use Phpml\Clustering\KMeans\Cluster; |
8
|
|
|
use Phpml\Clustering\KMeans\Point; |
9
|
|
|
use Phpml\Clustering\KMeans\Space; |
10
|
|
|
use Phpml\Exception\InvalidArgumentException; |
11
|
|
|
use Phpml\Math\Distance\Euclidean; |
12
|
|
|
|
13
|
|
|
class FuzzyCMeans implements Clusterer |
14
|
|
|
{ |
15
|
|
|
/** |
16
|
|
|
* @var int |
17
|
|
|
*/ |
18
|
|
|
private $clustersNumber; |
19
|
|
|
|
20
|
|
|
/** |
21
|
|
|
* @var Cluster[] |
22
|
|
|
*/ |
23
|
|
|
private $clusters = []; |
24
|
|
|
|
25
|
|
|
/** |
26
|
|
|
* @var Space |
27
|
|
|
*/ |
28
|
|
|
private $space; |
29
|
|
|
|
30
|
|
|
/** |
31
|
|
|
* @var float[][] |
32
|
|
|
*/ |
33
|
|
|
private $membership = []; |
34
|
|
|
|
35
|
|
|
/** |
36
|
|
|
* @var float |
37
|
|
|
*/ |
38
|
|
|
private $fuzziness; |
39
|
|
|
|
40
|
|
|
/** |
41
|
|
|
* @var float |
42
|
|
|
*/ |
43
|
|
|
private $epsilon; |
44
|
|
|
|
45
|
|
|
/** |
46
|
|
|
* @var int |
47
|
|
|
*/ |
48
|
|
|
private $maxIterations; |
49
|
|
|
|
50
|
|
|
/** |
51
|
|
|
* @var int |
52
|
|
|
*/ |
53
|
|
|
private $sampleCount; |
54
|
|
|
|
55
|
|
|
/** |
56
|
|
|
* @var array |
57
|
|
|
*/ |
58
|
|
|
private $samples = []; |
59
|
|
|
|
60
|
|
|
/** |
61
|
|
|
* @throws InvalidArgumentException |
62
|
|
|
*/ |
63
|
|
|
public function __construct(int $clustersNumber, float $fuzziness = 2.0, float $epsilon = 1e-2, int $maxIterations = 100) |
64
|
|
|
{ |
65
|
|
|
if ($clustersNumber <= 0) { |
66
|
|
|
throw new InvalidArgumentException('Invalid clusters number'); |
67
|
|
|
} |
68
|
|
|
|
69
|
|
|
$this->clustersNumber = $clustersNumber; |
70
|
|
|
$this->fuzziness = $fuzziness; |
71
|
|
|
$this->epsilon = $epsilon; |
72
|
|
|
$this->maxIterations = $maxIterations; |
73
|
|
|
} |
74
|
|
|
|
75
|
|
|
public function getMembershipMatrix(): array |
76
|
|
|
{ |
77
|
|
|
return $this->membership; |
78
|
|
|
} |
79
|
|
|
|
80
|
|
|
public function cluster(array $samples): array |
81
|
|
|
{ |
82
|
|
|
// Initialize variables, clusters and membership matrix |
83
|
|
|
$this->sampleCount = count($samples); |
84
|
|
|
$this->samples = &$samples; |
85
|
|
|
$this->space = new Space(count($samples[0])); |
86
|
|
|
$this->initClusters(); |
87
|
|
|
|
88
|
|
|
// Our goal is minimizing the objective value while |
89
|
|
|
// executing the clustering steps at a maximum number of iterations |
90
|
|
|
$lastObjective = 0.0; |
91
|
|
|
$iterations = 0; |
92
|
|
|
do { |
93
|
|
|
// Update the membership matrix and cluster centers, respectively |
94
|
|
|
$this->updateMembershipMatrix(); |
95
|
|
|
$this->updateClusters(); |
96
|
|
|
|
97
|
|
|
// Calculate the new value of the objective function |
98
|
|
|
$objectiveVal = $this->getObjective(); |
99
|
|
|
$difference = abs($lastObjective - $objectiveVal); |
100
|
|
|
$lastObjective = $objectiveVal; |
101
|
|
|
} while ($difference > $this->epsilon && $iterations++ <= $this->maxIterations); |
102
|
|
|
|
103
|
|
|
// Attach (hard cluster) each data point to the nearest cluster |
104
|
|
|
for ($k = 0; $k < $this->sampleCount; ++$k) { |
105
|
|
|
$column = array_column($this->membership, $k); |
106
|
|
|
arsort($column); |
107
|
|
|
reset($column); |
108
|
|
|
$cluster = $this->clusters[key($column)]; |
109
|
|
|
$cluster->attach(new Point($this->samples[$k])); |
110
|
|
|
} |
111
|
|
|
|
112
|
|
|
// Return grouped samples |
113
|
|
|
$grouped = []; |
114
|
|
|
foreach ($this->clusters as $cluster) { |
115
|
|
|
$grouped[] = $cluster->getPoints(); |
116
|
|
|
} |
117
|
|
|
|
118
|
|
|
return $grouped; |
119
|
|
|
} |
120
|
|
|
|
121
|
|
|
protected function initClusters(): void |
122
|
|
|
{ |
123
|
|
|
// Membership array is a matrix of cluster number by sample counts |
124
|
|
|
// We initilize the membership array with random values |
125
|
|
|
$dim = $this->space->getDimension(); |
126
|
|
|
$this->generateRandomMembership($dim, $this->sampleCount); |
127
|
|
|
$this->updateClusters(); |
128
|
|
|
} |
129
|
|
|
|
130
|
|
|
protected function generateRandomMembership(int $rows, int $cols): void |
131
|
|
|
{ |
132
|
|
|
$this->membership = []; |
133
|
|
|
for ($i = 0; $i < $rows; ++$i) { |
134
|
|
|
$row = []; |
135
|
|
|
$total = 0.0; |
136
|
|
|
for ($k = 0; $k < $cols; ++$k) { |
137
|
|
|
$val = random_int(1, 5) / 10.0; |
138
|
|
|
$row[] = $val; |
139
|
|
|
$total += $val; |
140
|
|
|
} |
141
|
|
|
|
142
|
|
|
$this->membership[] = array_map(static function ($val) use ($total): float { |
143
|
|
|
return $val / $total; |
144
|
|
|
}, $row); |
145
|
|
|
} |
146
|
|
|
} |
147
|
|
|
|
148
|
|
|
protected function updateClusters(): void |
149
|
|
|
{ |
150
|
|
|
$dim = $this->space->getDimension(); |
151
|
|
|
if (count($this->clusters) === 0) { |
152
|
|
|
for ($i = 0; $i < $this->clustersNumber; ++$i) { |
153
|
|
|
$this->clusters[] = new Cluster($this->space, array_fill(0, $dim, 0.0)); |
154
|
|
|
} |
155
|
|
|
} |
156
|
|
|
|
157
|
|
|
for ($i = 0; $i < $this->clustersNumber; ++$i) { |
158
|
|
|
$cluster = $this->clusters[$i]; |
159
|
|
|
$center = $cluster->getCoordinates(); |
160
|
|
|
for ($k = 0; $k < $dim; ++$k) { |
161
|
|
|
$a = $this->getMembershipRowTotal($i, $k, true); |
162
|
|
|
$b = $this->getMembershipRowTotal($i, $k, false); |
163
|
|
|
$center[$k] = $a / $b; |
164
|
|
|
} |
165
|
|
|
|
166
|
|
|
$cluster->setCoordinates($center); |
167
|
|
|
} |
168
|
|
|
} |
169
|
|
|
|
170
|
|
|
protected function getMembershipRowTotal(int $row, int $col, bool $multiply): float |
171
|
|
|
{ |
172
|
|
|
$sum = 0.0; |
173
|
|
|
for ($k = 0; $k < $this->sampleCount; ++$k) { |
174
|
|
|
$val = $this->membership[$row][$k] ** $this->fuzziness; |
175
|
|
|
if ($multiply) { |
176
|
|
|
$val *= $this->samples[$k][$col]; |
177
|
|
|
} |
178
|
|
|
|
179
|
|
|
$sum += $val; |
180
|
|
|
} |
181
|
|
|
|
182
|
|
|
return $sum; |
183
|
|
|
} |
184
|
|
|
|
185
|
|
|
protected function updateMembershipMatrix(): void |
186
|
|
|
{ |
187
|
|
|
for ($i = 0; $i < $this->clustersNumber; ++$i) { |
188
|
|
|
for ($k = 0; $k < $this->sampleCount; ++$k) { |
189
|
|
|
$distCalc = $this->getDistanceCalc($i, $k); |
190
|
|
|
$this->membership[$i][$k] = 1.0 / $distCalc; |
191
|
|
|
} |
192
|
|
|
} |
193
|
|
|
} |
194
|
|
|
|
195
|
|
|
protected function getDistanceCalc(int $row, int $col): float |
196
|
|
|
{ |
197
|
|
|
$sum = 0.0; |
198
|
|
|
$distance = new Euclidean(); |
199
|
|
|
$dist1 = $distance->distance( |
200
|
|
|
$this->clusters[$row]->getCoordinates(), |
201
|
|
|
$this->samples[$col] |
202
|
|
|
); |
203
|
|
|
|
204
|
|
|
for ($j = 0; $j < $this->clustersNumber; ++$j) { |
205
|
|
|
$dist2 = $distance->distance( |
206
|
|
|
$this->clusters[$j]->getCoordinates(), |
207
|
|
|
$this->samples[$col] |
208
|
|
|
); |
209
|
|
|
|
210
|
|
|
$val = (($dist1 / $dist2) ** 2.0) / ($this->fuzziness - 1); |
211
|
|
|
$sum += $val; |
212
|
|
|
} |
213
|
|
|
|
214
|
|
|
return $sum; |
215
|
|
|
} |
216
|
|
|
|
217
|
|
|
/** |
218
|
|
|
* The objective is to minimize the distance between all data points |
219
|
|
|
* and all cluster centers. This method returns the summation of all |
220
|
|
|
* these distances |
221
|
|
|
*/ |
222
|
|
|
protected function getObjective(): float |
223
|
|
|
{ |
224
|
|
|
$sum = 0.0; |
225
|
|
|
$distance = new Euclidean(); |
226
|
|
|
for ($i = 0; $i < $this->clustersNumber; ++$i) { |
227
|
|
|
$clust = $this->clusters[$i]->getCoordinates(); |
228
|
|
|
for ($k = 0; $k < $this->sampleCount; ++$k) { |
229
|
|
|
$point = $this->samples[$k]; |
230
|
|
|
$sum += $distance->distance($clust, $point); |
231
|
|
|
} |
232
|
|
|
} |
233
|
|
|
|
234
|
|
|
return $sum; |
235
|
|
|
} |
236
|
|
|
} |
237
|
|
|
|