Advancing Post Hoc Case Based Explanation with Feature Highlighting

Case-based explanations work better when they highlight specific features: The study found that showing users both exemplar cases AND highlighted features from the test image more effectively calibrated their confidence in AI predictions compared to showing cases alone, particularly for ambiguous classifications.

Feature highlighting helps users understand "why" beyond just "what": By isolating and connecting multiple clear feature parts of images to relevant training examples, the approach provides more comprehensive explanations that link back to the actual data the AI learned from, rather than treating explanations as a black box.

User studies are essential for validating XAI methods: The researchers emphasized that effectiveness of explanation techniques should be measured through carefully designed user studies rather than assumed, demonstrating that psychological validity matters when AI systems need to collaborate with humans.

This addresses a real gap in image-based AI explanations: Prior case-based explanation methods for images lacked the ability to show multiple relevant features and connect them to training data examples, leaving users with incomplete understanding of model decisions—this work fills that methodological gap.

Appropriate calibration of user confidence is a key goal: Rather than simply making users trust AI more, responsible explanation should help users develop accurate confidence levels in predictions, especially when classifications are genuinely ambiguous in real-world data.