Integrating Feature Understandability into Feature Selection for Human-Centred Tabular Explanations

Feature understandability matters for explanations, but implementation is complex: The research found that explanations using more understandable features scored higher on average, but the improvement was not statistically significant, suggesting that simply selecting easier-to-understand features alone may not be sufficient to meaningfully improve how well people comprehend AI explanations.

Current XAI evaluation happens too late in the process: The study highlights that most explainable AI research evaluates explanations only after they're generated, missing opportunities to improve accessibility from the start. This research proposes incorporating human feedback about feature understandability earlier—during the feature selection stage—to build better explanations from the ground up.

The Feature Understandability Scale provides a practical measurement tool: Researchers developed a method to quantify how well people understand different features in datasets by collecting user feedback and assigning numerical scores. This creates a concrete way to measure and compare feature complexity across different datasets and domains.

Designing human-centered AI requires iterative user input: The research demonstrates that improving explanations for real people requires collecting actual user feedback about what they find understandable, rather than relying solely on technical metrics. This underscores that responsible AI development must involve end-users throughout the design process, not just at the end.