Achieving Robust and Faithful Explanations via Conditional Pairwise Contrast

Explainable Artificial Intelligence (XAI) is crucial for understanding complex models like deep neural networks, thereby enabling trust, debugging, fairness assessment, and regulatory compliance, particularly in high-stakes domains such as healthcare and manufacturing. Local Interpretable Model-agnostic Explanations (LIME) stands out as one of the most widespread techniques in this category. In multi-class classification settings, standard LIME generates a single “one-vs-all” explanation that fails to capture the accurate decision logic between competing classes, making its explanations unstable and low fidelity. This paper introduces CondLIME, an extension framework that resolves these issues by fundamentally reframing the local explanation task. CondLIME generates a series of pairwise conditional explanations, with each linear model focusing on the distinction between the predicted class and a specific competing class. These conditional models are then innovatively used as candidate splitting rules to construct an oblique decision tree. This tree acts as a high-fidelity and complete explanation, defining a precise and robust classification region. Its hierarchy indicates that CondLIME provides explanations incrementally with a more precise scope. Our experiments demonstrate that CondLIME significantly outperforms standard LIME, delivering more stable explanations with substantially higher local fidelity. By providing a more reliable and multi-faceted insight into a model's behavior, CondLIME enhances the trustworthiness of local explanations for critical applications.

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.