Achieving descriptive accuracy in explanations via argumentation: The case of probabilistic classifiers

Area of Science:

• Artificial Intelligence
• Explainable AI (XAI)
• Human-Computer Interaction

Background:

• Trustworthy and fair AI systems are increasingly important for human-centric goals.
• Explanations are commonly used to build trust, but their properties require scrutiny.
• Descriptive accuracy (DA), the correspondence of explanations to AI internal workings, has been overlooked.

Purpose of the Study:

• To formalize descriptive accuracy (DA) in AI explanations.
• To analyze the satisfaction of DA notions by various explanation methods.
• To investigate the impact of DA on user trust and AI fairness.

Main Methods:

• Formal definitions of naive, structural, and dialectical DA were developed.
• Probabilistic classifiers were used as the analytical context.
• Several explanation methods, including feature-attribution techniques and a novel method, were evaluated for DA satisfaction.

Main Results:

• Formal definitions for different levels of DA were established.
• Experimental evaluation showed that existing methods may violate DA, particularly the demanding dialectical DA.
• A novel explanation method was proposed that satisfies dialectical DA by design.

Conclusions:

• Descriptive accuracy is a critical component for achieving trustworthy and fair AI systems.
• The proposed formalizations and novel method advance the field of Explainable AI.
• Prioritizing DA in explanations is essential for human-centric AI principles.