Source-linked AI summary
Explainability Fact Sheets: A Framework for Systematic Assessment of Explainable Approaches
Kacper Sokol, Peter Flach
TL;DR
The paper addresses the absence of consensus on which properties explainable systems should satisfy and develops a literature-based taxonomy spanning five dimensions. It operationalizes that taxonomy as Explainability Fact Sheets, enabling systematic comparison and disclosure of capabilities, limitations, and implementation discrepancies.
Problem
The rapid growth of XAI has made approaches difficult to compare because clearly defined evaluation properties and consensus on desired properties are lacking.
Method
The authors survey XAI literature, organize its criteria into five dimensions, and operationalize them as Explainability Fact Sheets and Work Sheets.
Results
The framework systematically evaluates and compares explainability approaches while identifying discrepancies between their theoretical properties and implementations.
Takeaways & Limitations
Researchers and practitioners can use Fact Sheets to report capabilities and limitations and guide the design, deployment, and evaluation of explainability methods.
Takeaways & Limitations
The requirements list is not final or definitive and is intended to be validated and revised through feedback from users and the XAI community.
Abstract
from arXiv · showhide
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.
1 INTRODUCTION
The paper addresses the difficulty of comparing rapidly proliferating XAI approaches without clearly defined evaluation properties. It proposes Explainability Fact Sheets and Work Sheets to make capabilities, limitations, and implementation discrepancies systematic and explicit.
- A lack of clearly defined evaluation properties hinders comparison and leaves limitations and implementation discrepancies undiscovered or undisclosed.
- Explainability Fact Sheets assess functional, operational, usability, safety, and validation dimensions.The safety dimension includes security, privacy, and vulnerabilities, while validation may use user studies or synthetic experiments.
- A standardized Fact Sheet can reveal properties lost between theoretical guarantees and implementations shaped by application domains or datasets.The paper gives counterfactual methods that may fail to exploit their social and interactive aspects as an example.
- The framework organizes characteristics from computer science and social science literature into a common basis for evaluating explainability approaches.
- Researchers and practitioners can use the template to discuss, evaluate, report, design, deploy, or assess explainability methods, including for best-practice or regulatory requirements.The paper demonstrates the practical framework with an instantiation for LIME.
2 EXPLAINABILITY FACT SHEETS DIMENSIONS
Explainability Fact Sheets classify methods by functional, operational, and deployment-related properties, while also documenting explanation forms, interactions, domains, and assumptions. These descriptors help users identify suitable approaches and expose requirements relevant to interpretation and use.
- Functional Requirements: Functional descriptors specify applicable learning tasks, problem types, explanation targets, explanation scope, computational complexity, model portability, feature types, and caveats.
- Deployment Perspective: These properties function as a deployment-oriented checklist for identifying explainability methods suitable for a particular use case.
- Functional Requirements: Explanation scope distinguishes local explanations for individual points, cohort explanations for subgroups, and global explanations of comprehensive model behavior.
- Functional Requirements: Model portability ranges from model-agnostic methods to model-class-specific and model-specific approaches.
- Functional Requirements: Fact Sheets should state compatible feature types, assumptions, and restrictions because preprocessing, correlated features, or input-output constraints can affect applicability.
- Operational Requirements: Operational descriptors cover explanation families, media, interaction protocols, and whether explanations use the original or a transformed domain.Media include statistical summaries, visualizations, textualizations, formal argumentation, or mixtures; interaction may be static or interactive.
3 EXPLAINABLE SYSTEMS TRADE-OFFS
Explainability requirements can conflict, vary by application, and be lost through implementation choices, so evaluations should make these trade-offs explicit.
- Functional and operational properties may conflict, and their importance often depends on the application area.
- Model-agnostic explainers add complexity and cannot exploit details of a particular model implementation, which can reduce fidelity.
- Vanilla counterfactual explanations prioritise completeness over soundness because they are always data point-specific.
- Social-science findings favour necessary, intentional, distinguishable, short, selective explanations shaped by social context.
- Explanation complexity can be adjusted through system design, user interaction, parsimony, or a hierarchy based on the user’s mental model.
4 DISCUSSION
The Fact Sheet framework is intended as a systematic, flexible reporting and development aid, while acknowledging adoption, validation, and completeness challenges.
- Systematic evaluation can precede user studies and reveal when implementation loses theoretically guaranteed explainability properties.
- Fact Sheets organise requirements across functional, operational, usability, safety, and validation dimensions for systematic reporting and comparison.
- Their unified format helps users identify implementation gaps, choose approaches for use cases, and communicate structured information to regulators and certification bodies.
- The evolving requirements list is intended for online versioning, dissemination, revision, and community-supported updates, while formal standards might slow adoption.
- Fact Sheets can be created incrementally or post-hoc, although the authors recommend using the template during development as a guideline.
- The requirements combine computational and social-science perspectives, with detailed categories retained for versatility across transparent algorithms and standalone techniques.
- Exhaustive Fact Sheets are labour- and time-consuming, and properties require either quantitative measurement or qualitative discussion and user studies.
5 RELATED WORK
Related work offers diverse explainability properties, evaluation practices, and reporting analogies, but does not provide the same organised evaluation framework.
- Explainability research draws on a longer humanities tradition, while AI literature presents desired properties across a scattered range of sources.
- Prior studies evaluate selected approaches using criteria such as fidelity, interaction, parsimony, actionability, soundness, and completeness.
- User studies are often treated as a gold standard, although some research questions whether they fully assess explanation effectiveness.
- Data statements, data sheets, nutrition labels, and privacy labels provide precedents for coherent self-reporting and certification.
6 CONCLUSIONS AND FUTURE WORK
The paper consolidates explainability requirements into Fact Sheets and begins extending the framework through implementation reviews of popular techniques.
- The paper covers functional, operational, usability, safety, and validation characteristics and uses them to assess explainable ML approaches and their trade-offs.
- Supplementary material provides an example Fact Sheet that closely inspects LIME across all five dimensions.
- Future work includes Fact Sheets for surrogates and counterfactuals, with implementation reviews examining theoretical and algorithmic capability discrepancies.
Local Interpretable
The Explainability Fact Sheet documents LIME, a surrogate method for interpreting individual predictions through a sparse linear approximation of a local decision boundary.
- Local Interpretable: The Fact Sheet is supplementary material accompanying the 2020 Explainability Fact Sheets paper.
- Local Interpretable: LIME is a surrogate explainability method for interpreting individual predictions.Its implementation can explain tabular, image, and text data.
- Local Interpretable: LIME approximates a local decision boundary with a sparse linear model.
bLIMEy
LIME is a post-hoc, model-agnostic local explainer whose pipeline transforms data, samples around an instance, queries the global model, weights samples by proximity, and fits local models. Its implementation supports multiple feature types and classifiers, but discretisation and un-discretisation introduce locality and stability caveats, while explanations remain instance-specific.
- Explanation Scope: LIME explanations are local and do not generalise beyond the data point for which they were generated.This limits their breadth to individual predictions rather than cohort- or global-level explanations.
- Computational Complexity: For each explained point, LIME may transform data, sample a neighbourhood, query the global model, compute distances, weight samples, select features, and train local class-versus-rest models.The required steps and their costs depend on the algorithmic components used.
- Model Relation: LIME is model agnostic and post-hoc, so it can work with any predictive model and be retrofitted to predictive systems.The official implementation nevertheless requires probabilistic black-box outputs for classification because it uses linear regression locally.
- Compatible Feature Types: LIME supports categorical and numerical features, while its default interpretable representation converts features into categorical bins for explanation legibility.Images use binary super-pixel vectors and text uses binary bag-of-words vectors.
- Caveats and Assumptions: Default tabular discretisation makes sampled data resemble a global rather than local neighbourhood before proximity weighting restores locality.The subsequent un-discretisation step uniformly samples within bins, introducing additional randomness.
- Usability Limitations: LIME explanations expose local feature importance but cannot quantify a factor’s effect to precisely guide future actions.They are also static, non-personalised, and cannot directly adjust explanation complexity.