Scholarlink

<p>The contribution of artificial intelligence (AI) to solving real-world problems is well established, but the issue of trusting AI systems remains an ongoing challenge. AI systems are considered to be complex because of their decision-making mechanism, which are not easily understandable by users. The stacking model is one of these complex models that is widely used because it offers better predictive performance. However, many studies focus solely on exploiting predictive performance without paying much attention to understanding the predictions of the stacking model. Nevertheless, among the few studies that have addressed this aspect of explainability, some have used existing agnostic methods to provide explanations, while others provide explanations based on a subset of base learners rather than the entire stacking model. In this work, a hybrid method is proposed and designed to explain the predictions of stacking models based on their internal behavior. The proposed method combines the Layer-wise Relevance Propagation (LRP) with model-agnostic feature importance techniques, such as SHapley Additive exPlanations (SHAP) or Local Interpretable Model-agnostic Explanations (LIME). Experimental results show that the method consistently exhibits greater stability than both SHAP and LIME, and in certain cases, provides higher fidelity relative to the size of the explanation.</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

Hybrid Method to Explain Predictions of Stacking Ensemble Model

  • 期刊论文

  • 发表日期：

    2026年1月7日

查看全文
  • Ingrid Pamela Nguemkam Tebou,
  • Norbert Tsopze,
  • Dieudonné Tchuente

摘要

The contribution of artificial intelligence (AI) to solving real-world problems is well established, but the issue of trusting AI systems remains an ongoing challenge. AI systems are considered to be complex because of their decision-making mechanism, which are not easily understandable by users. The stacking model is one of these complex models that is widely used because it offers better predictive performance. However, many studies focus solely on exploiting predictive performance without paying much attention to understanding the predictions of the stacking model. Nevertheless, among the few studies that have addressed this aspect of explainability, some have used existing agnostic methods to provide explanations, while others provide explanations based on a subset of base learners rather than the entire stacking model. In this work, a hybrid method is proposed and designed to explain the predictions of stacking models based on their internal behavior. The proposed method combines the Layer-wise Relevance Propagation (LRP) with model-agnostic feature importance techniques, such as SHapley Additive exPlanations (SHAP) or Local Interpretable Model-agnostic Explanations (LIME). Experimental results show that the method consistently exhibits greater stability than both SHAP and LIME, and in certain cases, provides higher fidelity relative to the size of the explanation.