Trusted-AI/AIX360

Interpretability and explainability of data and machine learning models

What it solves

AI Explainability 360 (AIX360) addresses the "black box" problem in machine learning by providing a comprehensive set of tools to make the decisions of AI models and the nature of their datasets more interpretable and transparent. It recognizes that no single explanation method works for every scenario, offering a variety of algorithms to suit different data types and use cases.

How it works

The toolkit provides a library of explainability algorithms categorized by the type of explanation they provide:

  • Data Explanations: Tools to explain the underlying datasets.
  • Local Post-hoc Explanations: Methods to explain individual predictions after a model has been trained (e.g., LIME, SHAP).
  • Time-Series Explanations: Specialized local post-hoc tools for time-series data, including saliency maps and adapted LIME.
  • Local Direct Explanations: Methods where the model is designed to be inherently explainable at a local level.
  • Global Direct Explanations: Models that are inherently interpretable across the whole dataset (e.g., Boolean Decision Rules).
  • Global Post-hoc Explanations: Methods to explain the overall behavior of a trained model.
  • Certification: Tools like Ecertify to probabilistically certify local explanations.

It also includes proxy metrics such as Faithfulness and Monotonicity to evaluate the quality of the explanations produced.

Who it’s for

AIX360 is designed for data scientists and AI researchers who need to implement interpretability techniques into their ML pipelines, as well as consumers who want a gentle introduction to AI explainability concepts via the provided interactive experience.

Highlights

  • Multi-modal Support: Works with tabular, text, image, and time-series data.
  • Comprehensive Taxonomy: Includes a wide array of algorithms ranging from local to global and direct to post-hoc explanations.
  • Extensible Design: Built to allow contributors to add new algorithms and metrics.
  • Evaluation Metrics: Provides built-in metrics to measure the faithfulness and monotonicity of explanations.

Related

  • Project
    microsoft/responsible-ai-toolbox

    A suite of model and data exploration tools that enable developers to identify errors, assess fairness, and interpret AI model predictions for more responsible deployment.

  • Project
    Trusted-AI/AIF360

    An extensible open-source library for Python and R that provides metrics and algorithms to detect and mitigate algorithmic bias in machine learning models.

  • Project
    interpretml/interpret

    An open-source framework that provides both inherently interpretable glassbox models and tools to explain complex blackbox machine learning systems.

  • Project
    deel-ai/xplique

    A Python toolkit for Explainable AI (XAI) that provides tools for attribution, feature visualization, and concept extraction to help users understand complex neural network models.

  • Project
    MAIF/shapash

    A Python library that makes machine learning models interpretable for everyone by providing user-friendly visualizations, an interactive web app, and audit reports.