Release v0.1.3.post1: Enhanced Statistical Analysis & Developer Experience

Enhanced Statistical Analysis & Developer Experience

• IV Standard Errors: Add confidence intervals and significance testing for Information Value calculations
• Series Input Support: Fix 'Series' object has no attribute 'columns' error
• Type Checking Integration: Integrate pyrefly type checking with CI workflows
• FAISS GPU Support: Add fastwoe[faiss-gpu] optional dependency and improve dependency management
• Improved UX: Change default warn_on_numerical=False (numerical binning is now core feature)
• Local Testing: Add act integration for local GitHub Actions testing
• Enhanced Documentation: Comprehensive guides and examples with improved mathematical framework
• Development Automation: Add Make commands for streamlined workflows

Release v0.1.2.post4

Changes

• See CHANGELOG.md for detailed changes

Installation

pip install fastwoe==v0.1.2.post4

Release v0.1.2.post3

Changes

• See CHANGELOG.md for detailed changes

Installation

pip install fastwoe==v0.1.2.post3

v0.1.2

Version 0.1.2

• Improved:
  • Added decision tree-based binning (DecisionTreeClassifier) for numerical features
  • Implemented fast Somers' D in numba for non-binary target variables (e.g., loss rates)
    • Fastest Somers' D implementation in Python (3x faster than scipy)
    • Produces both D_Y|X and D_X|Y scores

v0.1.1post4

v0.1.1.post4

Fixed:

• sklearn Compatibility: Simplified sklearn compatibility by always using quantile_method="averaged_inverted_cdf" parameter since FastWoe requires scikit-learn>=1.3.0 where this parameter is always supported.

Technical Details:

• Removed unnecessary sklearn version detection code
• Always applies quantile_method="averaged_inverted_cdf" for consistent binning behavior
• Maintains scikit-learn>=1.3.0 requirement

Notes:

• All changes from v0.1.1.post1, v0.1.1.post2, and v0.1.1.post3 are included
• This release supersedes all previous post releases

v0.1.1post3

v0.1.1.post3

Fixed:

• sklearn Version Compatibility: Resolved TypeError with KBinsDiscretizer for users with sklearn versions < 1.3.0. The quantile_method parameter is now only used when supported by the installed sklearn version, with graceful fallback for older versions.
• API Consistency: predict_ci() method now returns a numpy array with shape (n_samples, 2) containing [ci_lower, ci_upper] instead of a DataFrame, making it consistent with predict_proba() which returns (n_samples, 2) for [prob_class_0, prob_class_1].
• Enhanced Input Support: All prediction methods (predict(), predict_proba(), predict_ci()) now accept numpy arrays as input, automatically converting them to pandas DataFrames with warnings for consistency.

Technical Details:

• Added sklearn version detection to conditionally use quantile_method="averaged_inverted_cdf" parameter
• Updated WeightOfEvidence.predict_ci() in interpret module to match new API
• Comprehensive test coverage for sklearn version compatibility

Notes:

• All changes from v0.1.1.post1 and v0.1.1.post2 are included in this release
• This release supersedes v0.1.1.post1 and v0.1.1.post2
• Breaking change: predict_ci() now returns numpy array instead of DataFrame

v0.1.1post2

v0.1.1.post2

• Fixed:
  • NumPy array input handling: FastWoe.fit and related methods now accept NumPy arrays as input, automatically converting them to pandas DataFrames/Series with a warning. This prevents AttributeError and improves user experience.
• Notes:
  • All changes from 0.1.1.post1 are included in this release.
  • This release supersedes 0.1.1.post1.

v0.1.1.post1

• Bug Fixes:
  • Fixed issues with pandas/numpy data type conversions
  • Improved handling of rare categories in WOE calculations
  • Better error messages for edge cases

v0.1.1

🚀 Release v0.1.1

✨ New Features

• WeightOfEvidence Interpretability Module - Complete explanation framework for FastWoe classifiers
• Auto-Inference Capabilities - Automatically detect feature names, class names, and training data
• Confidence Intervals - Uncertainty quantification for WOE scores and predictions
• Enhanced API - Dual usage patterns with explain(sample) and explain(dataset, index)

📊 Technical Improvements

• 95 comprehensive tests with 92% code coverage
• Enhanced FastWoe core with confidence interval support
• Comprehensive documentation and working examples
• Cross-compatibility testing

📦 Changes

• 14 files changed: 4,109 insertions, 380 deletions
• 3 new major files: interpretability module, tests, and examples
• Updated dependencies and build configuration

v0.1.0 - Initial Beta Release

FastWoe v0.1.0 - Initial Beta Release

Fast Weight of Evidence (WOE) encoding and inference

✨ Features

• Core WOE Implementation: Fast encoding using scikit-learn's TargetEncoder
• Statistical Rigor: MLE-based standard errors and confidence intervals
• High-Cardinality Support: WoePreprocessor for handling many categories
• Cross-Version Compatibility: Python 3.9-3.12, sklearn 1.3.0+
• Complete Documentation: Mathematical background and examples

📋 Requirements

• Python 3.9+
• scikit-learn 1.3.0+
• pandas, numpy, scipy

🚀 Installation

pip install git+https://github.com/xRiskLab/fastwoe.git@v0.1.0

Note

This is a beta release. The API may change before v1.0.0.