The classical machine learning library.
WHAT IT PROVIDES
Consistent interfaces across many algorithms Preprocessing transformers Pipelines combining them Model selection and cross-validation Metrics
WHAT THE CONSISTENT INTERFACE MEANS
Fit on training data, then transform or predict.
WHY THAT MATTERS
Algorithms can be swapped without rewriting surrounding code.
WHAT A PIPELINE DOES
Chains preprocessing and a model into one object.
WHY IT IS ESSENTIAL
It prevents leakage, since preprocessing is fitted only on training folds.
WHAT HAPPENS WITHOUT ONE
Scaling fitted on all data, leaking test information.
WHAT COLUMN TRANSFORMERS PROVIDE
Different preprocessing per column type.
WHAT CROSS-VALIDATION UTILITIES PROVIDE
Correct splitting, including grouped and time-based variants.
WHAT TO USE FOR TIME-ORDERED DATA
The time-series split, never the default.
WHAT TO BE CAREFUL WITH
Default parameters, which are starting points not recommendations Metrics chosen by default Fitting transformers outside a pipeline
WHAT IT DOES NOT COVER WELL
Deep learning Very large datasets Gradient boosting at the level dedicated libraries provide
WHAT TO PAIR IT WITH
A dedicated boosting library, for tabular problems.
