1
|
|
|
<?php |
2
|
|
|
|
3
|
|
|
declare(strict_types=1); |
4
|
|
|
|
5
|
|
|
namespace Phpml\DimensionReduction; |
6
|
|
|
|
7
|
|
|
use Phpml\Exception\InvalidArgumentException; |
8
|
|
|
use Phpml\Exception\InvalidOperationException; |
9
|
|
|
use Phpml\Math\Statistic\Covariance; |
10
|
|
|
use Phpml\Math\Statistic\Mean; |
11
|
|
|
|
12
|
|
|
class PCA extends EigenTransformerBase |
13
|
|
|
{ |
14
|
|
|
/** |
15
|
|
|
* Temporary storage for mean values for each dimension in given data |
16
|
|
|
* |
17
|
|
|
* @var array |
18
|
|
|
*/ |
19
|
|
|
protected $means = []; |
20
|
|
|
|
21
|
|
|
/** |
22
|
|
|
* @var bool |
23
|
|
|
*/ |
24
|
|
|
protected $fit = false; |
25
|
|
|
|
26
|
|
|
/** |
27
|
|
|
* PCA (Principal Component Analysis) used to explain given |
28
|
|
|
* data with lower number of dimensions. This analysis transforms the |
29
|
|
|
* data to a lower dimensional version of it by conserving a proportion of total variance |
30
|
|
|
* within the data. It is a lossy data compression technique.<br> |
31
|
|
|
* |
32
|
|
|
* @param float $totalVariance Total explained variance to be preserved |
33
|
|
|
* @param int $numFeatures Number of features to be preserved |
34
|
|
|
* |
35
|
|
|
* @throws InvalidArgumentException |
36
|
|
|
*/ |
37
|
|
|
public function __construct(?float $totalVariance = null, ?int $numFeatures = null) |
38
|
|
|
{ |
39
|
|
|
if ($totalVariance !== null && ($totalVariance < 0.1 || $totalVariance > 0.99)) { |
40
|
|
|
throw new InvalidArgumentException('Total variance can be a value between 0.1 and 0.99'); |
41
|
|
|
} |
42
|
|
|
|
43
|
|
|
if ($numFeatures !== null && $numFeatures <= 0) { |
44
|
|
|
throw new InvalidArgumentException('Number of features to be preserved should be greater than 0'); |
45
|
|
|
} |
46
|
|
|
|
47
|
|
|
if (($totalVariance !== null) === ($numFeatures !== null)) { |
48
|
|
|
throw new InvalidArgumentException('Either totalVariance or numFeatures should be specified in order to run the algorithm'); |
49
|
|
|
} |
50
|
|
|
|
51
|
|
|
if ($numFeatures !== null) { |
52
|
|
|
$this->numFeatures = $numFeatures; |
53
|
|
|
} |
54
|
|
|
|
55
|
|
|
if ($totalVariance !== null) { |
56
|
|
|
$this->totalVariance = $totalVariance; |
57
|
|
|
} |
58
|
|
|
} |
59
|
|
|
|
60
|
|
|
/** |
61
|
|
|
* Takes a data and returns a lower dimensional version |
62
|
|
|
* of this data while preserving $totalVariance or $numFeatures. <br> |
63
|
|
|
* $data is an n-by-m matrix and returned array is |
64
|
|
|
* n-by-k matrix where k <= m |
65
|
|
|
*/ |
66
|
|
|
public function fit(array $data): array |
67
|
|
|
{ |
68
|
|
|
$n = count($data[0]); |
69
|
|
|
|
70
|
|
|
$data = $this->normalize($data, $n); |
71
|
|
|
|
72
|
|
|
$covMatrix = Covariance::covarianceMatrix($data, array_fill(0, $n, 0)); |
73
|
|
|
|
74
|
|
|
$this->eigenDecomposition($covMatrix); |
75
|
|
|
|
76
|
|
|
$this->fit = true; |
77
|
|
|
|
78
|
|
|
return $this->reduce($data); |
79
|
|
|
} |
80
|
|
|
|
81
|
|
|
/** |
82
|
|
|
* Transforms the given sample to a lower dimensional vector by using |
83
|
|
|
* the eigenVectors obtained in the last run of <code>fit</code>. |
84
|
|
|
* |
85
|
|
|
* @throws InvalidOperationException |
86
|
|
|
*/ |
87
|
|
|
public function transform(array $sample): array |
88
|
|
|
{ |
89
|
|
|
if (!$this->fit) { |
90
|
|
|
throw new InvalidOperationException('PCA has not been fitted with respect to original dataset, please run PCA::fit() first'); |
91
|
|
|
} |
92
|
|
|
|
93
|
|
|
if (!is_array($sample[0])) { |
94
|
|
|
$sample = [$sample]; |
95
|
|
|
} |
96
|
|
|
|
97
|
|
|
$sample = $this->normalize($sample, count($sample[0])); |
98
|
|
|
|
99
|
|
|
return $this->reduce($sample); |
100
|
|
|
} |
101
|
|
|
|
102
|
|
|
protected function calculateMeans(array $data, int $n): void |
103
|
|
|
{ |
104
|
|
|
// Calculate means for each dimension |
105
|
|
|
$this->means = []; |
106
|
|
|
for ($i = 0; $i < $n; ++$i) { |
107
|
|
|
$column = array_column($data, $i); |
108
|
|
|
$this->means[] = Mean::arithmetic($column); |
109
|
|
|
} |
110
|
|
|
} |
111
|
|
|
|
112
|
|
|
/** |
113
|
|
|
* Normalization of the data includes subtracting mean from |
114
|
|
|
* each dimension therefore dimensions will be centered to zero |
115
|
|
|
*/ |
116
|
|
|
protected function normalize(array $data, int $n): array |
117
|
|
|
{ |
118
|
|
|
if (count($this->means) === 0) { |
119
|
|
|
$this->calculateMeans($data, $n); |
120
|
|
|
} |
121
|
|
|
|
122
|
|
|
// Normalize data |
123
|
|
|
foreach (array_keys($data) as $i) { |
124
|
|
|
for ($k = 0; $k < $n; ++$k) { |
125
|
|
|
$data[$i][$k] -= $this->means[$k]; |
126
|
|
|
} |
127
|
|
|
} |
128
|
|
|
|
129
|
|
|
return $data; |
130
|
|
|
} |
131
|
|
|
} |
132
|
|
|
|