Implementation with PHP ?..

Data transformation is essential for machine learning, as it prepares raw data into a format suitable for analysis and modeling. This chapter explores four key transformation techniques using RubixML and PHP-ML: Encoding Categorical Variables, Normalizing and Scaling Numerical Features, Reshaping Data Structures, and Feature Engineering.

---

RubixML Examples

1. Encoding Categorical Variables

Categorical data, such as "color" or "size," needs to be converted into numerical format so machine learning models can interpret it. One-Hot Encoding is a common method that transforms each category into a binary vector.

Example with RubixML:

use Rubix\ML\Datasets\Unlabeled;
use Rubix\ML\Transformers\OneHotEncoder;

// Create the dataset
$dataset = new Unlabeled([
    ['red', 'small'],
    ['blue', 'medium'],
    ['green', 'large'],
]);

$encoder = new OneHotEncoder();
$encoder->fit($dataset);
$samples = $dataset->samples();
$encoder->transform($samples);

echo "\nAfter Encoding:\n";
foreach ($samples as $sample) {
    echo implode('', $sample) . "\n";
}

Here, OneHotEncoder from RubixML converts each unique category into binary values, making it compatible with machine learning algorithms.

After Encoding:
100100
010010
001001

2. Normalizing and Scaling Numerical Features

Normalization adjusts numerical data to a standard range (often [0, 1]), which helps with model performance when features are on different scales.

Example with RubixML:

use Rubix\ML\Datasets\Labeled;
use Rubix\ML\Transformers\MinMaxNormalizer;

// Create the dataset
$dataset = new Labeled([
    [2000, 300],
    [2400, 450],
    [3000, 500],
], ['low', 'medium', 'high']);

$normalizer = new MinMaxNormalizer();
$normalizer->fit($dataset);

$samples = $dataset->samples();
$labels = $dataset->labels();
$normalizer->transform($samples);

echo "\nNormalized data:\n";
foreach ($samples as $ind => $sample) {
    echo implode(',', $sample) . ',' . $labels[$ind] . "\n";
}

In this example, MinMaxNormalizer scales values to the [0, 1] range, ensuring each feature is comparable.

Normalized data:
0,0,low
0.4,0.75,medium
1,1,high

3. Reshaping Data Structures

Reshaping allows us to organize data into structures required by specific algorithms. For example, in time series analysis, data can be reshaped into rolling windows for sequence modeling.

Example of Reshaping for Time Series:

$data = [[100], [150], [200], [250], [300], [350]];

// Create windows manually since RubixML doesn't have built-in windowing
function reshapeIntoRollingWindows(array $data, int $windowSize): array {
    // If input is a flat array, convert each element to an array
    $isFlat = !is_array(reset($data));
    $formattedData = $isFlat ? array_map(fn($value) => [$value], $data) : $data;

    $windows = [];
    for ($i = 0; $i <= count($formattedData) - $windowSize; $i++) {
        $window = array_slice($formattedData, $i, $windowSize);
        $windows[] = array_column($window, 0);
    }
    return $windows;
}

$reshapedData = reshapeIntoRollingWindows($dataset->samples(), 3);

// Convert back to RubixML dataset if needed
$windowedDataset = new Unlabeled($reshapedData);

In this example, reshapeIntoRollingWindows manually reshaping the dataset into sequences of three-day periods, making it suitable for time series models.

After Reshaping: 
[[100, 150, 200], [150, 200, 250], [200, 250, 300], [250, 300, 350]]

4. Feature Engineering

Feature engineering enhances model performance by creating new attributes from existing data. For instance, polynomial expansion generates interaction terms between features, which can reveal complex patterns.

Example with RubixML:

use Rubix\ML\Transformers\PolynomialExpander;

$dataset = new Labeled([
    [2000, 300],
    [2500, 400],
    [3000, 500],
], ['low', 'medium', 'high']);

$expander = new PolynomialExpander(2); // Creates second-degree polynomial features
$expander->transform($dataset);

print_r($dataset->samples());

The PolynomialExpander in RubixML generates interaction terms for each feature pair, allowing the model to capture non-linear relationships between attributes.

---

PHP-ML Examples

1. Encoding Categorical Variables

PHP-ML also provides one-hot encoding for categorical data, which is crucial for converting non-numerical values into binary format.

Example with PHP-ML:

use Phpml\Preprocessing\OneHotEncoder;

$samples = [
    ['red', 'small'],
    ['blue', 'medium'],
    ['green', 'large'],
];

$encoder = new OneHotEncoder();
$encoder->fit($samples);
$encoder->transform($samples);

print_r($samples);

OneHotEncoder in PHP-ML performs the same categorical transformation, making categorical values accessible to models.

2. Normalizing and Scaling Numerical Features

PHP-ML provides a Normalizer class to scale numerical values, ensuring that all features are on a comparable scale.

Example with PHP-ML:

use Phpml\Preprocessing\Normalizer;

$samples = [
    [2000, 300],
    [2500, 400],
    [3000, 500],
];

$normalizer = new Normalizer();
$normalizer->transform($samples);

print_r($samples);

Here, the Normalizer class in PHP-ML scales each feature within the dataset, which can significantly enhance compatibility with algorithms sensitive to differing feature scales.

3. Reshaping Data Structures

While PHP-ML does not have a built-in reshaping function, reshaping data can be done manually to suit sequence or time series modeling requirements.

Example of Reshaping for Time Series:

$data = [100, 150, 200, 250, 300, 350];
$windowSize = 3;
$reshapedData = [];

for ($i = 0; $i <= count($data) - $windowSize; $i++) {
    $reshapedData[] = array_slice($data, $i, $windowSize);
}

print_r($reshapedData);

This custom reshaping structure groups data into rolling windows, allowing it to be used for models that require sequential data.

4. Feature Engineering

PHP-ML does not include specific feature engineering tools like polynomial expansion, but custom features can be added manually to enhance model performance.

Manual Feature Engineering Example in PHP-ML:

$samples = [
    [2000, 300],
    [2500, 400],
    [3000, 500],
];

foreach ($samples as &$sample) {
    $sample[] = $sample[0] * $sample[1]; // Creating an interaction term between the first and second feature
}

print_r($samples);

In this example, we manually add a new feature by calculating the product of two existing features, introducing a potential interaction term that may help the model identify more complex patterns.

---

Summary

Data transformation prepares raw data for analysis by encoding, normalizing, reshaping, and engineering features. Using RubixML and PHP-ML, these transformations can be efficiently implemented in PHP, enhancing data compatibility and model accuracy. In the next chapter, we will explore feature selection, discussing ways to retain the most relevant features for improving model efficiency and accuracy.

To try this code yourself, install the example files from the official GitHub repository: https://github.com/apphp/ai-with-php-examples