Skip to content

CHANGELOG.md

Latest commit

 

History

History
183 lines (147 loc) · 8.06 KB

CHANGELOG.md

File metadata and controls

183 lines (147 loc) · 8.06 KB

Changelog

[0.2.7] - 2025-12-04

Changed

  • CatBoost Naming Standardization: Replaced LeafValue with XAddEvidence throughout CatBoost implementation
    • Standardized naming to match XGBoost and LightGBM implementations
    • Updated all CatBoost-related files: catboost_scorecard.py, catboost_wrapper.py, cb_constructor.py
    • Updated all tests and documentation

Fixed

  • README Documentation: Corrected CatBoost depth requirement statement
    • Changed from "Only supports depth=1" to "depth=1 is recommended for better interpretability"
    • Code actually supports any tree depth (as long as trees are complete binary)
    • Updated code examples to use XAddEvidence instead of LeafValue

Added

  • PyPI Publish Workflow: Added automated PyPI publishing workflow (.github/workflows/publish.yml)
    • Supports both release events and manual workflow dispatch
    • Uses trusted publishing (OpenID Connect) for secure PyPI uploads
    • Automatically uploads distribution files to GitHub releases

Technical Details

  • All 106 tests passing
  • Version updated from 0.2.7rc2 to 0.2.7 (stable release)
  • LightGBM support is now stable (previously release candidate)

[0.2.7rc2] - 2025-11-23 (Release Candidate)

Fixed

  • Complete Implementation: Added missing create_points() and predict_score() methods
    • Previous rc1 release had incomplete squash merge
    • Now includes all critical bug fixes and implementations

[0.2.7rc1] - 2025-11-23 (Release Candidate - Incomplete)

Added

  • LightGBM Scorecard Support: Complete implementation of LGBScorecardConstructor
    • Implemented create_points() with proper base_score normalization
    • Implemented predict_score() and predict_scores() for scorecard predictions
    • Added use_base_score parameter for flexible base score handling
    • Full parity with XGBoost scorecard functionality

Fixed

  • Critical Bug Fix: Corrected leaf ID mapping in extract_leaf_weights()
    • Changed from cumcount() to extracting actual leaf ID from node_index string
    • Fixes 55% Gini loss (0.40 → 0.90) in scorecard predictions
    • Ensures correct mapping between LightGBM's absolute leaf IDs and relative indices
  • Base Score Normalization: Proper handling of LightGBM's base score
    • Subtract base_score from Tree 0 leaves to balance tree contributions
    • Add logit(base_score) during scaling to distribute across all trees
    • Prevents first tree from getting disproportional weight

Changed

  • Simplified Score Types: Only XAddEvidence supported for LightGBM
    • Removed WOE support (ill-defined for LightGBM's sklearn API)
    • Cleaner, more maintainable implementation
  • Enhanced Documentation: Updated docstrings and examples
    • Added comprehensive LightGBM getting-started notebook
    • Explained base_score handling differences from XGBoost

Technical Details

  • All 106 tests passing (9 LightGBM-specific tests)
  • Scorecard Gini: 0.9020 vs Model Gini: 0.9021 (perfect preservation)
  • Proper handling of LightGBM's sklearn API vs internal booster API
  • Related to PR #8

[0.2.7a2] - 2025-11-08 (Alpha)

Added

  • LightGBM Support (Alpha): Initial implementation of LGBScorecardConstructor for LightGBM models
    • Implemented extract_leaf_weights() method for parsing LightGBM tree structure
    • Implemented get_leafs() method for leaf indices and margins prediction
    • Added comprehensive test suite with 9 tests (all passing)
    • Created example demonstrating implemented functionality
    • Status: Alpha release - 2 of 5 core methods implemented
    • Community: Reference implementation for issue #7 (@RektPunk)

Changed

  • Development Workflow: Consolidated shell scripts into enhanced Makefile
    • Removed run-ci-checks.sh and run-act.sh scripts
    • Enhanced Makefile with colored output and Docker checks
    • Added new targets: make ci-check, make act-local, make test-quick
    • Single entry point for all development tasks
    • make ci-check runs fast local checks (no Docker)
    • make act-* commands use Docker for GitHub Actions simulation

Fixed

  • CI/CD: Added missing lightgbm>=4.0.0,<5.0.0 dependency
  • Type Checking: Added LightGBM type stubs and configured ignore rules
  • GitHub Actions: Fixed act configuration by removing unsupported flags

Technical Details

  • LightGBM constructor follows same pattern as XGBoost/CatBoost implementations
  • Column naming unified to XAddEvidence for consistency across all constructors
  • Test suite expanded from 95 to 104 tests
  • All tests passing in local and CI environments

Installation

# Install latest alpha release
pip install git+https://github.com/xRiskLab/xBooster.git@v0.2.7a2

[0.2.7] - 2025-11-06

Changed

  • Code Optimization: Simplified get_leafs() and construct_scorecard() methods in XGBScorecardConstructor (PR #6)
    • Removed special-case branching for first iteration
    • Precomputes full leaf index matrix once instead of repeated predictions
    • Eliminates redundant DataFrame concatenations
    • Net reduction of 40 lines of code while maintaining identical functionality

Added

  • Comprehensive Regression Tests: Added 13 new tests to verify code refactoring produces identical outputs
  • Build System Improvements: Modernized hatchling configuration
    • Simplified version management in __init__.py
    • Removed setuptools legacy configuration
    • Added explicit egg-info exclusion
  • Code Quality Tools: Added prek and ty type checker configurations
  • Type Stubs Directory: Created typings/ directory for custom type definitions

Fixed

  • Improved .gitignore to properly exclude build artifacts and egg-info files

Technical Details

  • All 95 tests passing with no regressions
  • Leaf indices now stored as float32 (XGBoost's default) but represent whole numbers
  • Float precision differences negligible (< 1e-6)
  • Performance maintained across all operations

[0.2.6.post1] - 2025-09-30

Changed

  • XGBoost Compatibility: Extended dependency range from >=2.0.0,<3.0.0 to >=2.0.0,<4.0.0
  • Test Precision: Updated test_extract_model_param to handle XGBoost 3.0.5 precision differences
  • CI/CD Enhancement: Added comprehensive XGBoost version matrix testing (2.1.4, 3.0.5, latest)

Added

  • New Test Suite: Added test_xgboost_compatibility.py with 8 comprehensive compatibility tests
  • Enhanced Workflows: Updated GitHub Actions to test across multiple XGBoost versions
  • Better Error Handling: Improved Pylint configuration for virtual environment compatibility

Compatibility

  • ✅ Verified compatibility with XGBoost 3.0.5
  • ✅ Backward compatible with XGBoost 2.x versions
  • ✅ All existing functionality remains unchanged
  • ✅ No breaking changes for existing users

Technical Details

  • Fixed precision differences in base_score parameter extraction between XGBoost versions
  • Enhanced CI pipeline to catch compatibility issues early
  • Improved development environment setup with better Pylint integration

[0.2.6] - 2025-08-30

  • Added interval scorecard functionality for XGBoost models with max_depth=1
  • New methods: construct_scorecard_by_intervals() and create_points_peo_pdo()
  • Simplifies complex tree rules into interpretable intervals following industry standards (Siddiqi, 2017)
  • Typically achieves 60-80% rule reduction while maintaining accuracy

[0.2.5] - 2025-04-19

  • Minor changes in catboost_wrapper.py and cb_constructor.py to improve the scorecard generation.

[0.2.4] - 2025-04-18

  • Changed the build distribution in pyproject.toml.

[0.2.3] - 2025-04-18

  • Added support for CatBoost classification models and switch to uv for packaging.
  • Python version requirement updated to 3.10-3.11.

[0.2.2] - 2024-05-08

  • Updates in explainer.py module to improve kwargs handling and minor changes.

[0.2.1] - 2024-05-03

  • Updates of dependencies

[0.2.0] - 2024-05-03

  • Added tree visualization class (explainer.py)
  • Updated the local explanation algorithm for models with a depth > 1 (explainer.py)
  • Added a categorical preprocessor (_utils.py)

[0.1.0] - 2024-02-14

  • Initial release