Sometimes, one or more preprocessing steps will need to be taken to condition your data for a learner. Some examples of preprocessing include feature extraction, standardization, normalization, imputation, and dimensionality reduction. Preprocessing in Rubix ML is handled through Transformer objects whose logic is hidden behind an easy-to-use interface. Each transformer performs a pass over the samples in a dataset and alters the features in some way. Stateful transformers are fitted to a dataset before transforming. After fitting a transformer, it will expect the features to be present in the same order when transforming subsequent datasets.
Transform a Dataset#
One example of a transformation involves converting the categorical features of a dataset to continuous ones using a one hot encoding. To accomplish this with the library, pass a One Hot Encoder instance to a Dataset object's
apply() method which automatically handles fitting the dataset and transforming the samples.
use Rubix\ML\Transformers\OneHotEncoder; $dataset->apply(new OneHotEncoder());
Transformations can be chained by calling the
apply() method fluently.
use Rubix\ML\Transformers\RandomHotDeckImputer; use Rubix\ML\Transformers\OneHotEncoder; use Rubix\ML\Transformers\ZScaleStandardizer; $dataset->apply(new RandomHotDeckImputer(5)) ->apply(new OneHotEncoder()) ->apply(new ZScaleStandardizer());
Pipeline meta-estimators help you automate a series of transformations. In addition, Pipeline objects are Persistable which allow you to save and load transformer fittings between processes. Whenever a dataset object is passed to a learner wrapped in a Pipeline, it will transparently be fitted and/or transformed in the background before it arrives in the method context.
Let's apply the same 3 transformers as in the example above by passing the transformer instances in the order we want them applied along with a base estimator to the constructor of Pipeline like in the example below.
use Rubix\ML\Pipeline; use Rubix\ML\Transformers\RandomHotDeckImputer; use Rubix\ML\Transformers\OneHotEncoder; use Rubix\ML\Transformers\ZScaleStandardizer; use Rubix\ML\Classifiers\SoftmaxClassifier; $estimator = new Pipeline([ new RandomHotDeckImputer(5), new OneHotEncoder(), new ZScaleStandardizer(), ], new SoftmaxClassifier(200));
partial() will result in the transformers being fitted or updated before being passed to the Softmax Classifier.
$estimator->train($dataset); // Transformers fitted and applied automatically $estimator->partial($dataset); // Transformers updated and applied
Any time a dataset is passed to the Pipeline it will automatically be transformed before being handed to the underlying estimator.
$predictions = $estimator->predict($dataset); // Dataset automatically transformed
Standardization and Normalization#
Oftentimes, the continuous features of a dataset will be on different scales because they were measured by different methods. For example, age (0 - 100) and income (0 - 9,999,999) are on two widely different scales. Standardization is the processes of transforming a dataset such that the features are all on one scale. Normalization is the special case where the transformed features have a range between 0 and 1. Depending on the transformer, it may operate on the columns or the rows of the dataset.
|L1 Normalizer||Rows||[0, 1]|
|L2 Normalizer||Rows||[0, 1]|
|Max Absolute Scaler||Columns||[-1, 1]||●||●|
|Min Max Normalizer||Columns||[min, max]||●||●|
|Robust Standardizer||Columns||[-∞, ∞]||●|
|Z Scale Standardizer||Columns||[-∞, ∞]||●||●|
Since learners can be compatible with different data types, it may be necessary to convert features of an incompatible type to a compatible one. Feature converters do just that for all the columns of a dataset in a fast and efficient manner.
|One Hot Encoder||Categorical||Continuous||●|
|Numeric String Converter||Categorical||Continuous|
One technique for handling missing data is a preprocessing step called imputation. Imputation is the process of replacing missing values in the dataset with a pretty good substitution. Examples include the average value for a feature or the sample's nearest neighbor's value. Imputation allows you to get more value from your data and limits the introduction of certain biases in the process.
|KNN Imputer||Continuous, Categorical||●|
|Missing Data Imputer||Continuous, Categorical||●|
|Random Hot Deck Imputer||Continuous, Categorical||●|
Higher-order data such as images and text blobs are actually composites of many scalar features. Thus, it is often necessary to extract those features from their original representation in order to train a learner.
|Word Count Vectorizer||Text Blobs||●|
Dimensionality reduction in machine learning is analogous to compression in the context of sending data over a wire. It allows a learner to train and infer quicker by producing a dataset with fewer but more informative features.
|Dense Random Projector||●|
|Gaussian Random Projector||●|
|Linear Discriminant Analysis||●||●|
|Principal Component Analysis||●|
|Sparse Random Projector||●|
Similarly to dimensionality reduction, feature selection aims to reduce the number of features in a dataset, however, feature selection seeks to keep the best features as-is and drop the less informative ones entirely. Adding feature selection can help speed up training and inference by creating a more parsimonious model. It can also improve the performance of the model by removing noise features and features that are uncorrelated with the outcome.
For natural language processing (NLP) tasks, cleaning the text will help eliminate noise such as stop words or other uninformative tokens like URLs and email addresses from the corpus. Another common step is to normalize the text so that words like
ThErApIsT are recognized as the same word.
For computer vision tasks, images may need to be processed to ensure they are the correct size and shape. Other forms of image processing may include color correction and blurring/sharpening.