Understanding preferences for explainable AI: A study on LIME variations.

The current research aims to empirically examine user preferences of four variations of visual explanations from an image classification model. Participants (N = 23) were recruited from a Midwestern university. The study was a within-subjects design in which participants were asked to rate four local interpretable model-agnostic explanation (LIME) variations from best to worst in an image classification task across 10 trials. A Friedman's test and a Plackett-Luce analysis were conducted to examine participants' rank-order preferences for the variations, and a linear mixed-effects model on eye-tracking data was used to examine on which variations participants expended the most visual effort. Results reveal that, overall, participants preferred images that were non-obfuscated and spent more time fixating on images that highlighted both positive and negative features compared to positive only features. The findings have implications for interface design, particularly for designers aiming to prioritize human-centered preferences.