PREF-XAI: Preference-Based Personalized Rule Explanations of Black-Box Machine Learning Models

One-size-fits-all explanations are insufficient: The paper demonstrates that different users need different explanations based on their goals, preferences, and cognitive abilities. Traditional AI explanations that simply approximate how a model works often fail to meet individual user needs, highlighting a gap between current XAI practice and what users actually require.

User preferences can be learned efficiently from minimal feedback: PREF-XAI shows that by asking users to rank just a small set of candidate explanations, the system can accurately learn their individual preferences and reconstruct what types of explanations matter most to them—suggesting that personalization doesn't require extensive user input.

Personalized explanations can reveal insights users didn't anticipate: The research found that preference-based explanations can "discover novel explanatory rules not initially considered by the user," meaning that tailoring explanations to individual preferences may help users understand AI decisions in ways they hadn't previously thought to ask about.

Formal preference modeling provides a principled alternative to guesswork: Rather than relying on heuristics or implicit assumptions about what users want, PREF-XAI uses robust ordinal regression to formally model preferences—offering a more systematic and transparent approach to personalization that can be audited and improved.