Declarative Reasoning on Explanations Using Constraint Logic Programming

Explanations can incorporate real-world knowledge: REASONX allows users to add background knowledge and common sense reasoning to AI explanations through linear constraints, addressing a key limitation where current explanation methods often ignore important context that humans use to understand decisions.

Users should be able to interact with explanations at different levels of detail: The system enables "constraint projection," allowing users to explore explanations at varying levels of abstraction rather than receiving a single fixed explanation, making AI decisions more understandable to different audiences.

Explanations work better with simpler, interpretable models: REASONX focuses on decision trees as either the primary model or as surrogate models for black-box systems, suggesting that building explanations on top of inherently interpretable models may be more effective than trying to explain opaque systems directly.

Declarative reasoning provides a transparent explanation framework: By using Constraint Logic Programming and Prolog, REASONX grounds explanations in formal logic with clear semantics, potentially making the explanation process itself more auditable and trustworthy compared to explanation methods based on less transparent computational approaches.

Contrastive explanations (showing what would need to change) enhance understanding: The method uses both factual instances and contrastive scenarios to help users understand not just why a decision was made, but what factors would need to differ to produce a different outcome.