Skip to content

Latest commit

 

History

History
69 lines (52 loc) · 5.22 KB

CHANGELOG.md

File metadata and controls

69 lines (52 loc) · 5.22 KB

Changelog

All notable changes to MCGrad will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

Unreleased

Fixed

  • Fixed documentation inaccuracies in the MCE metrics page: corrected mce formatting from percent to absolute scale, added documentation for MDE and all four available scales, and clarified statistical interpretation of sigma threshold and p-value.

Changed

  • Improved early stopping performance by caching per-fold metric DataFrames and fold splits, avoiding redundant DataFrame reconstruction on every round × fold × metric evaluation (#229)
  • Speed up fitting with early stopping and save training performance enabled: Remove redundant _predict call on the training set in _determine_best_num_rounds, reusing predictions already returned by _fit_single_round instead (#236)
  • Unified logistic() and logistic_vectorized() into a single logistic() function that handles both scalar and array inputs. The logistic_vectorized alias has been removed. Performance improvement of ~16x on large arrays by using native NumPy operations instead of np.vectorize.
  • Vectorized segments_ecce_pvalue computation in MulticalibrationError using NumPy broadcasting instead of np.vectorize. Code quality improvement; does not meaningfully affect end-to-end MCE computation time since p-value calculation is <0.1% of the pipeline (dominated by segment generation and ECCE computation).
  • Eliminated redundant segments * sample_weight broadcast in _calculate_cumulative_differences (~6% faster on the ECCE computation hot path).
  • Eliminated redundant segments * sample_weight broadcast in _ecce_standard_deviation_per_segment (~30% faster on the std dev computation hot path).
  • Replaced per-segment pd.DataFrame creation with batch list accumulation in get_segment_masks, yielding 2.2-3.1x speedup on the segment generation hot path (~53% of total pipeline time).

Added

  • Conda / Mamba installation instructions to the documentation (#234)
  • youdens_j metric (Youden's J statistic / informedness) in the metrics module. Computes the continuous net detection rate as E[scores | positive] - E[scores | negative], ranging from -1 to 1, with support for sample weights (#233)
  • Support for float calibration targets (soft labels) in MCGrad. Labels can now be continuous values in [0, 1] instead of only binary 0/1 or True/False. This enables calibrating to confidence-weighted labels that account for label uncertainty and quality.
  • soft_label_log_loss function in the metrics module for soft-label cross-entropy, used as the default early stopping metric in MCGrad.
  • Feature consistency check in predict() that validates feature names and order match those used during fit(). Previously, mismatched features produced confusing encoder errors or silently wrong predictions when feature order was swapped.
  • Feature names are now included in the serialized model JSON for consistency validation after deserialization
  • Regression support in MulticalibrationError — new outcome_type="regression" parameter uses successive-differences variance estimation (equation 6, appendix C from the MCE paper) instead of Bernoulli variance (#218)
  • Regression support in tune_mcgrad_params — now accepts RegressionMCGrad models, uses the model's early_stopping_score_func instead of hardcoded normalized_entropy (#213)

[0.1.3] - 2026-02-11

Changed

  • Lowered psutil dependency to 5.9.0 to fix compatibility with Google Colab (#210)

[0.1.2] - 2026-02-10

Fixed

  • Fixed optimization direction in tune_mcgrad_params — was incorrectly maximizing normalized entropy instead of minimizing it (#202)

Added

  • use_model_predictions parameter to tune_mcgrad_params to control whether the surrogate model's predicted best or the actual best observed trial is returned (#196)
  • mcgrad.tutorials submodule with shared tutorial helpers (data loading, plotting, logging utilities)

Changed

  • Moved folktables from optional [tutorials] dependency to core dependencies

[0.1.1] - 2026-01-30

Note: Early pre-release versions (i.e., 0.0.0 and 0.1.0) were yanked due to packaging/versioning cleanup; 0.1.1 is the first stable public release.

Added

  • Initial release of this library. To cite this library, please cite our KDD 2026 paper "MCGrad: Multicalibration at Web Scale"
  • Core multicalibration algorithm (MCGrad, RegressionMCGrad)
  • Traditional calibration methods (IsotonicRegression, PlattScaling, TemperatureScaling)
  • Segment-aware calibrators for grouped calibration
  • Comprehensive metrics module with calibration error metrics, Kuiper statistics, and ranking metrics
  • MulticalibrationError class for measuring multicalibration quality
  • Plotting utilities for calibration curves and diagnostics
  • Hyperparameter tuning integration with Ax platform
  • Full documentation website at mcgrad.dev
  • Sphinx API documentation at mcgrad.readthedocs.io