Changelog¶

All notable changes to this project are documented here.

The format follows Keep a Changelog. This project adheres to Semantic Versioning.

[0.3.4] — 2026-07-09¶

Declared Python support as 3.9 through 3.12 in package metadata and documentation.

transform() now detects data that appears to have already been transformed by the same fitted transformer, returns it unchanged, and emits a warning.

Parallel per-column processing via the n_jobs constructor argument for fit() and transform().
Support for joblib-style negative n_jobs values, including n_jobs=-1 to use all available CPUs.

inverse_transform() now converts datetime columns from their integer offset representation back to pandas timestamps.
Datetime missing-value restoration now uses NaT, preserving datetime dtype when restore_missing=True.

Category mapping accessors: get_category_mappings() and get_category_mapping(...).
Portable fitted-state persistence with save() and IFCTransformer.load() for transform/inverse-transform workflows on another machine.

The default categorical missing category is now __ifcfill_missing__ to reduce collisions with real categories.
Label encoding now runs as a separate layer after categorical filling so learned fill categories are included in the mapping.
Categorical columns with one observed category plus missing values are kept when cat_fill="constant" because the missing category is learnable.

Optional categorical label encoding via cat_encoding="label".
Category mapping attributes for inverse-compatible label encoding: category_mappings_ and inverse_category_mappings_.
Categorical missing values are represented as a learnable category by default and converted back to missing values during inverse_transform().

IFCTransformer class with fit(), transform(), fit_transform(), and inverse_transform().
Automatic type inference for integer, float, categorical, and datetime columns.
Per-column type overrides via col_types constructor argument.
Configurable imputation strategies:
Integer: mean, median, mode, zero
Float: mean, median, mode, zero
Categorical: mode, constant
Datetime-to-integer conversion with configurable anchor date and time unit.
Automatic detection and removal of constant columns.
Missing value distribution tracking via missing_counts_, missing_fractions_, and missing_report_.
inverse_transform() with optional missing-value restoration and reproducible random_state.
Accepts pandas.DataFrame and CSV file paths as input.
NumPy-based computation engine for fast processing.