Vfrdtyky

can someone help me with the difference in feature scaling using normalize vs normalizer as in

• preprocessing.normalize() vs preprocessing.Normalizer() .

They are equivalent in terms of their effect on the data. The normalize function is intended to be a 'quick and easy' option to normalise a single vector/matrix. Normalizer is what's known as a 'utility class'. It just wraps the normalize function in Sklearn's Transformer API. As stated in the documentation, this makes the Normalizer class well-suited for use in Sklearn's Pipeline class.

** Update to provide more details **

Normalisers

The functionality of normalize and Normalizer is identical. I.e. given the same data and parameters they will each return the L1 or L2 norm of the input matrix.

See the sklearn.preprocessing.normalizer documentation for more details.

Transformers

Normalizer is an example of a transformer. Transformers can process data in a wide variety of ways. One thing they all have in common in sklearn is having fit, transform and fit_transform methods.

See the data transformer documentation for more details.

Pipelines

sklearn.pipeline.Pipeline is a class for 'chaining' data transformations (normalisation, scaling, filtering etc.) and an estimator. Being able to do this is helpful when using cross-validation to optimise different parameters in the preprocessing transformations and the estimator.

Because of the way sklearn.pipeline.Pipeline works, it requires constituent functions to use the transformer API. I.e to have fit, transform and fit_transform methods.

See the sklearn.pipeline.Pipline documentation for more details.

Difference Between Normalize and Normalizer

Normalize does not have fit, transform and fit_transform methods. So whilst it is suited to 'standalone' use, it can't be used as part of a Pipeline. Normalizer is wrapped in sklearn's Transformer API in order to provide the Transformer methods.

Therefore Normalizer allows the normalize function to be used with the fit, transform and fit_transform methods, which in turn allows it to be used as part of a Pipeline.

Notes

1. Because there is no model to fit when normalising data (technically because, normalize is stateless apart from the configuration parameters, e.g. which axis to normalise on and which type of normalisation to use) it only needs the transform method. So fit returns the input without modification and fit_transform behaves the same as transform.


2. There may be other reasons for having the Transformer API available to normalize but its use as part of a Pipeline is the most common.