Evaluating XAI Methods for Static and Evolving Data in DetoxAI
A new paper from arXiv addresses the limitations of Explainable Artificial Intelligence (XAI) evaluation, using the DetoxAI image recognition system as a case study. The paper, accepted for publication in the proceedings of the EASi 2026 Workshop at IJCAI-ECAI 2026 in Bremen, explores human-grounded evaluation of explanation methods for image classification and discusses adapting counterfactual explanations to evolving data streams with concept drift. It also tackles the challenge of tracking the co-evolution of data, models, and explanations. The work is part of the broader field of XAI and is published in Springer CCIS vol 3107 (2016).
Key facts
- Paper discusses limitations of XAI evaluation.
- Uses DetoxAI image recognition system for bias detection and concept unlearning.
- Presents human-grounded evaluation of explanation methods for image classification.
- Explores adapting explanations to evolving data streams with concept drift.
- Discusses experiences with adapting counterfactuals.
- Relates to challenges of tracking co-evolution of data, models, and explanations.
- Accepted for publication in EASi 2026 Workshop at IJCAI-ECAI 2026 Bremen.
- Published in Springer CCIS vol 3107 (2016).
Entities
Institutions
- arXiv
- DetoxAI
- IJCAI-ECAI
- Springer
Locations
- Bremen