One Explanation Does Not Fit All: A Toolkit and Taxonomy of AI Explainability Techniques (1909.03012v2)

Published 6 Sep 2019 in cs.AI, cs.CV, cs.HC, and stat.ML

Abstract: As artificial intelligence and machine learning algorithms make further inroads into society, calls are increasing from multiple stakeholders for these algorithms to explain their outputs. At the same time, these stakeholders, whether they be affected citizens, government regulators, domain experts, or system developers, present different requirements for explanations. Toward addressing these needs, we introduce AI Explainability 360 (http://aix360.mybluemix.net/), an open-source software toolkit featuring eight diverse and state-of-the-art explainability methods and two evaluation metrics. Equally important, we provide a taxonomy to help entities requiring explanations to navigate the space of explanation methods, not only those in the toolkit but also in the broader literature on explainability. For data scientists and other users of the toolkit, we have implemented an extensible software architecture that organizes methods according to their place in the AI modeling pipeline. We also discuss enhancements to bring research innovations closer to consumers of explanations, ranging from simplified, more accessible versions of algorithms, to tutorials and an interactive web demo to introduce AI explainability to different audiences and application domains. Together, our toolkit and taxonomy can help identify gaps where more explainability methods are needed and provide a platform to incorporate them as they are developed.

Authors (20)

Vijay Arya (9 papers)
Rachel K. E. Bellamy (9 papers)
Pin-Yu Chen (311 papers)
Amit Dhurandhar (62 papers)
Michael Hind (25 papers)
Samuel C. Hoffman (13 papers)
Stephanie Houde (18 papers)
Q. Vera Liao (49 papers)
Ronny Luss (27 papers)
Sami Mourad (3 papers)
Pablo Pedemonte (3 papers)
Ramya Raghavendra (5 papers)
John Richards (16 papers)
Prasanna Sattigeri (70 papers)
Karthikeyan Shanmugam (85 papers)
Moninder Singh (17 papers)
Kush R. Varshney (121 papers)
Dennis Wei (64 papers)
Yunfeng Zhang (45 papers)
Aleksandra Mojsilović (5 papers)

Citations (357)

View on Semantic Scholar

Summary

Overview of "One Explanation Does Not Fit All: A Toolkit and Taxonomy of AI Explainability Techniques"

The paper "One Explanation Does Not Fit All: A Toolkit and Taxonomy of AI Explainability Techniques" introduces AI Explainability 360 (AIX360), a comprehensive open-source toolkit developed by researchers at IBM. This toolkit aims to address the diverse and growing need for explainability in AI systems, essential for building trust and meeting regulatory requirements. The paper outlines the architectural design, functionality, and underlying taxonomy of explainability methods, detailing how AIX360 facilitates better understanding and transparency.

Key Contributions

Taxonomy of Explainability Techniques:
- Comprehensive Framework: Proposes a structured taxonomy to categorize explainability techniques based on various criteria, such as level (local vs. global), timing (pre-, in-, post-model), and user type (e.g., data scientists, end-users).
- Real-World Relevance: Highlights how different stakeholders require distinct forms of explanations tailored to their needs.
AI Explainability 360 Toolkit:
- Diverse Methods: AIX360 includes eight state-of-the-art explainability methods, such as BRCG, GLRM, and CEM, allowing users to choose suitable techniques for specific applications.
- Access and Extensibility: Encourages community involvement by being open-source, providing APIs for easy integration and extension.
- Educational Material: Supplies tutorials and an interactive web demo to make explainability concepts accessible to non-experts.
Algorithmic Innovations and Enhancements:
- Simplified Implementations: Introduces a “light” version of BRCG to avoid reliance on proprietary solvers, making it more accessible.
- Data Synthesis for Training: Describes methods for creating synthetic datasets with explanations, enhancing applicability in practical settings.
Evaluation Metrics:
- Includes quantitative metrics like Faithfulness and Monotonicity to assess the quality of explanations, a feature not commonly found in other toolkits.

Implications

Practical Applications: The development of AIX360 paves the way for more transparent AI systems in sectors such as finance, healthcare, and human resources, contributing to user trust and compliance with regulations.
Research Opportunities: By identifying gaps and creating a taxonomy, the paper highlights areas for future exploration in interactive and data explainability methods, pushing the research agenda forward.
Community Engagement: The open-source nature of AIX360 promotes collaboration, allowing researchers and practitioners to contribute new methods and insights.

Future Directions

Expansion of Taxonomy: As the field of explainability evolves, expanding the taxonomy to include new methods and more intricate criteria remains a key area of development.
Interactive Explanations: While current methods are largely static, advancing towards interactive explanations could significantly enhance user understanding and satisfaction.
Framework Agnosticism: Developing framework-independent models could increase the usability of explainability methods across different AI systems.

Conclusion

The paper provides an important contribution by systematically organizing the complex landscape of AI explainability through AIX360. This toolkit, with its extensible architecture and thorough documentation, aims to democratize the use of explainability methods across various industries and research domains. As AI continues to integrate into critical decision-making processes, the need for tailored, comprehensive explainable AI solutions becomes increasingly crucial.

PDF Markdown

Related Papers

Find Related Papers

GitHub

GitHub - Trusted-AI/AIX360: Interpretability and explainability of data and machine learning models (1,702 stars)