Scholarlink

<p>The disagreement problem arises when multiple Explainable Artificial Intelligence (XAI) methods provide divergent explanations for the same machine learning model prediction. This inconsistency leaves practitioners uncertain about the reliability of post-hoc XAI methods and limits their adoption in high-stakes applications. In this work, we introduce Consensus Importance, a novel metric that analyzes explanations from multiple XAI methods using faithfulness-driven weighting methods to enhance explanation consistency. Our approach integrates regional clustering to identify homogeneous regions, quantified by high Overall Agreement (OA) scores, where explanations are not likely to disagree. We define quantitative agreement metrics that are used to measure consistency between LIME and SHAP. To support robustness, we present an uncertainty-aware diagnosis using argumentation frameworks. Our findings is based on ten diverse datasets with varying sizes and dimensions and multiple models, demonstrating that regional analysis combined with these strategies improves the reliability and interpretability of XAI outputs. This work contributes a more dependable explainability framework and paves a pathway for future research.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Quantifying overall agreement between explanation methods to address the disagreement problem

  • 期刊论文

  • 发表日期：

    2026年2月2日

查看全文
  • Gaurav Parashar,
  • Alka Chaudhary,
  • Dilkeshwar Pandey

摘要

The disagreement problem arises when multiple Explainable Artificial Intelligence (XAI) methods provide divergent explanations for the same machine learning model prediction. This inconsistency leaves practitioners uncertain about the reliability of post-hoc XAI methods and limits their adoption in high-stakes applications. In this work, we introduce Consensus Importance, a novel metric that analyzes explanations from multiple XAI methods using faithfulness-driven weighting methods to enhance explanation consistency. Our approach integrates regional clustering to identify homogeneous regions, quantified by high Overall Agreement (OA) scores, where explanations are not likely to disagree. We define quantitative agreement metrics that are used to measure consistency between LIME and SHAP. To support robustness, we present an uncertainty-aware diagnosis using argumentation frameworks. Our findings is based on ten diverse datasets with varying sizes and dimensions and multiple models, demonstrating that regional analysis combined with these strategies improves the reliability and interpretability of XAI outputs. This work contributes a more dependable explainability framework and paves a pathway for future research.