My collection of custom-made transformers based on Scikit-Learn pipelines

1. BASIC FEATURES MANIPULATIONS

• SimpleFeaturesTransformer - add new columns with prefixes cat__ and num__ for specified columns, that will be treated as initial features on the next steps of pipeline

• FeatureEliminationTransformer - drops features that are constant or quasi-constant overall dataset, and so are useless. Detects and drops duplicated features. Based on Feature-engine lib

• DecorrelationTransformer - detects and removes correlated features. Based on Feature-engine lib

  • Detects groups of correlated features based on correlation threshold set up
  • From every group only one feature is selected based on mean target performance. Other features within this group will be dropped

2. ENCODERS

• WoeEncoderTransformer - transforms categirical features into weight of evidence (WOE) values (see detailed explonation here). Based on Feature-engine lib

• RareCategoriesTransformer - encodes rare values of categorical features into one value Others to reduce features cardinality. Based on Feature-engine lib

• BinningCategoriesTransformer - applies categories grouping for categorical features into bigger groups with similar WOE values to reduce cardinality. Based on optbinning lib.

• BinningNumericalTransformer - applies discretisation for numerical features. Based on optbinning lib.

3. MODELING

• CustomLogisticRegressionClassifier - fits Logistic Regression model (from statsmodels), performs iterative feature selection excluding features with high p-values after every model fit

• CustomCatBoostClassifier - fits CatBoostClassifier model, performs iterative feature selection excluding features with low feature importance after every model fit. Uses evaluation set for early stopping.