Release Notes#
Version 0.1.7#
Introduced StandardComplementaryNeymanShapExplainer, a new explainer that utilizes standard Neyman allocation for SHAP value estimation. Existing ComplementaryNeymanShapExplainer has been renamed to LimitedComplementaryNeymanShapExplainer to reflect its limited sampling strategy.
Internal refactoring and bug fixes to improve code maintainability and performance.
Version 0.1.6#
Includes fixes for complementary-based explainers to ensure correct SHAP value computations.
Improved documentation and added more comprehensive tests for all modules.
Small fixes for bugs discovered in previous versions during production use.
Ensures minimal fraction is respected in HierarchicalExplainer when using importance sampling.
Version 0.1.5#
Added ComplementaryNeymanShapExplainer, a new explainer leveraging Neyman allocation for more efficient sampling.
Refactored base classes to improve code clarity, maintainability, and facilitate future extensions.
Fixed minor bugs in ComplementaryShapApproximation to ensure accurate SHAP value calculations.
Fixes bug in PreciseShapExplainer related to generator dead-lock.
Introduced new modes in HierarchicalExplainer for more flexible explanation strategies, including: - Multi-modal explanations. - Business-aware first-level splits (at the cost of performance). - Importance-aware approximation budgets (for each group) for better resource allocation.
Version 0.1.4#
Added ComplementaryShapApproximation, an explainer using complementary contribution sampling for faster SHAP value estimation.
Enhanced global tests to cover new functionalities and improve robustness.
Version 0.1.3#
Added HierarchicalExplainer, an explainer using hierarchical sampling to accelerate SHAP value estimation for long text inputs (text modality only).
Refactored base classes to improve code organization and maintainability.
Version 0.1.2#
Added Monte Carlo-inspired explainers: - LimitedMcShapExplainer: uses limited sampling strategies to efficiently estimate SHAP values. - LimitedMcShapExplainer (restricted to unique masks): improves sampling efficiency by avoiding duplicates.
Fixed issues in the connectors module to enhance stability and performance.
Version 0.1.1#
Initial release of MLLM-SHAP, a library for SHAP value estimation in Multi-Modal Large Language Models (MLLMs).
Added PreciseShapExplainer for exact SHAP value computation.
Introduced the connectors module with support for the Liquid Audio LMF-2 multi-modal model.