SelectionPipeline

class lightautoml.pipelines.selection.base.SelectionPipeline(features_pipeline=None, ml_algo=None, imp_estimator=None, fit_on_holdout=False, **kwargs)[source]

Bases: object

Abstract class, performing feature selection. Instance should accept train/valid datasets and select features.

property is_fitted

Check if selection pipeline is already fitted.

Return type

bool

Returns

True for fitted pipeline and False for not fitted.

property selected_features

Get selected features.

Return type

List[str]

Returns

List of selected feature names.

property in_features

Input features to the selector.

Raises exception if not fitted beforehand.

Return type

List[str]

Returns

List of input features.

property dropped_features

Features that were dropped.

Return type

List[str]

Returns

list of dropped features.

__init__(features_pipeline=None, ml_algo=None, imp_estimator=None, fit_on_holdout=False, **kwargs)[source]

Create features selection pipeline.

Parameters
perform_selection(train_valid)[source]

Select features from train-valid iterator.

Method is used to perform selection based on features pipeline and ml model. Should save _selected_features attribute in the end of working.

Raises

NotImplementedError.

fit(train_valid)[source]

Selection pipeline fit.

Find features selection for given dataset based on features pipeline and ml model.

Parameters

train_valid (TrainValidIterator) – Dataset iterator.

select(dataset)[source]

Takes only selected features from giving dataset and creates new dataset.

Parameters

dataset (LAMLDataset) – Dataset for feature selection.

Return type

LAMLDataset

Returns

New dataset with selected features only.

map_raw_feature_importances(raw_importances)[source]

Calculate input feature importances. Calculated as sum of importances on different levels of pipeline.

Parameters

raw_importances (Series) – Importances of output features.

get_features_score()[source]

Get input feature importances.

Returns

Series with importances in not ascending order.