Trusted-AI/AIX360
Interpretability and explainability of data and machine learning models
What it solves
AI Explainability 360 (AIX360) provides a comprehensive toolkit for making machine learning models and datasets more interpretable. Because no single explainability method works for every scenario, the library offers a wide variety of algorithms to help users find the right approach for their specific data type and use case.
How it works
The toolkit implements a diverse set of algorithms categorized by the type of explanation they provide:
- Data Explanations: Focuses on describing the underlying data itself.
- Local Post-hoc Explanations: Explains individual predictions after a model has been trained (e.g., LIME, SHAP).
- Time-Series Explanations: Specialized local post-hoc methods for time-series data.
- Local Direct Explanations: Provides explanations directly from the model's internal logic during prediction.
- Global Direct Explanations: Uses models that are inherently interpretable by design (e.g., Boolean Decision Rules).
- Global Post-hoc Explanations: Explains the overall behavior of a model globally after training.
- Certification: Includes tools to certify the reliability of local explanations.
It also includes proxy metrics like Faithfulness and Monotonicity to evaluate the quality of of the explanations produced.
Who it’s for
It is designed for data scientists and AI practitioners who need to understand, trust, and explain the decisions made by their machine learning models across tabular, text, image, and time-series data.
Highlights
- Multi-modal support: Works with tabular, text, images, and time-series data.
- Extensive algorithm library: Includes a broad range of techniques from LIME and SHAP to specialized time-series saliency maps.
- Taxonomy-based guidance: Provides a taxonomy tree and guidance material to help users select the appropriate algorithm.
- Evaluation metrics: Includes built-in metrics to measure the faithfulness and monotonicity of explanations.