Explainability Fact Sheets: A Framework for Systematic Assessment of Explainable Approaches (1912.05100v1)

Abstract: Explanations in Machine Learning come in many forms, but a consensus regarding their desired properties is yet to emerge. In this paper we introduce a taxonomy and a set of descriptors that can be used to characterise and systematically assess explainable systems along five key dimensions: functional, operational, usability, safety and validation. In order to design a comprehensive and representative taxonomy and associated descriptors we surveyed the eXplainable Artificial Intelligence literature, extracting the criteria and desiderata that other authors have proposed or implicitly used in their research. The survey includes papers introducing new explainability algorithms to see what criteria are used to guide their development and how these algorithms are evaluated, as well as papers proposing such criteria from both computer science and social science perspectives. This novel framework allows to systematically compare and contrast explainability approaches, not just to better understand their capabilities but also to identify discrepancies between their theoretical qualities and properties of their implementations. We developed an operationalisation of the framework in the form of Explainability Fact Sheets, which enable researchers and practitioners alike to quickly grasp capabilities and limitations of a particular explainable method. When used as a Work Sheet, our taxonomy can guide the development of new explainability approaches by aiding in their critical evaluation along the five proposed dimensions.

A Structured Framework for Explainable AI: An In-Depth Evaluation

The paper "Explainability Fact Sheets: A Framework for Systematic Assessment of Explainable Approaches" by Kacper Sokol and Peter Flach, introduces a comprehensive taxonomy and operational framework for assessing explainability methods in machine learning. This work aims to address the absence of a unified standard to evaluate explainable systems, which has been a barrier in the field of eXplainable Artificial Intelligence (XAI). The authors propose an Explainability Fact Sheet, a tool designed to systematically characterize and evaluate explainability approaches along five dimensions: functional, operational, usability, safety, and validation.

Taxonomy and Purpose

The authors performed an extensive survey of the literature related to explainable AI, focusing on both emerging algorithms and established criteria, to inform their taxonomy. This survey allowed them to extract key desiderata necessary for building a robust framework capable of assessing not only the theoretical qualities of explainable methods but also their practical implementations. The proposed taxonomy serves as a structured guide to catalog the capabilities and limitations of an explainability approach, benefiting both researchers and practitioners.

Core Dimensions

1. Functional Requirements: This dimension evaluates how an explainability method suits certain AI problems, touching upon factors such as problem type, applicable model classes, and computational complexity. For instance, whether a method is model-agnostic or specific to particular model families is critical for its applicability.
2. Operational Requirements: This aspect covers how the method interacts with end-users and its operational characteristics, like the medium of explanations and system interaction types. It gauges the balance between explainability and predictive performance, crucial for real-world deployment.
3. Usability Requirements: Perhaps the most nuanced, this dimension attends to the user-centered aspects, ensuring that explanations are comprehensible, actionable, and tailored to the needs of the audience. Properties like soundness, completeness, coherence, and parsimony are pivotal for fostering trust and reliability in AI systems.
4. Safety Requirements: Explainability methods must mitigate risks relating to privacy, security, and robustness. This involves measuring how much information an explanation reveals about the model and data, and the potential for adversarial misuse.
5. Validation Requirements: This dimension underscores the importance of empirically validating explainability methods, either through synthetic experiments or user studies. Verification processes ascertain the method's effectiveness and faithfulness to the theoretical underpinnings it claims to satisfy.

Implications and Future Directions

The introduction of Explainability Fact Sheets provides a structured medium for discussing, evaluating, and reporting the properties of explainable AI techniques. By unifying evaluation methods, these fact sheets promote transparency, comparability, and a higher standard of scrutiny in the design and deployment of XAI methods.

In practical terms, the framework's adoption could improve adherence to best practices and aid in compliance with regulations like the GDPR's "right to explanation." The methodical assessment this framework enables is not just beneficial for developers but also regulatory bodies and certification entities, ensuring AI models’ fairness and accountability.

Looking forward, this framework may evolve through community contributions and adaptations, fostering a culture of transparency in AI research. The prospect of hosting these Explainability Fact Sheets within a centralized online repository could facilitate ongoing refinement and widespread adoption, ultimately advancing the broader field of interpretable and transparent AI. Future work could explore measuring trade-offs between competing desiderata, as understanding these balances is crucial for the practical deployment of explainable systems.