Scholarlink

<p>The application of machine learning (ML) in slope stability assessment has been hindered by a lack of attention to uncertainty quantification and interpretability, which limits its adoption in engineering design. Prior studies have emphasized predictive accuracy without addressing the reliability and transparency necessary for decision-making in safety-critical contexts. This study introduces a novel framework for circular mode slope stability assessment that integrates split conformal prediction (SCP) with explainable ML to achieve statistically valid and interpretable predictions. The framework employs a Random Forest model optimized with the Artificial Hummingbird Algorithm (AHA) and implements SCP to generate calibrated prediction sets with guaranteed confidence. The dataset is divided into training, calibration, and testing subsets, where nonconformity scores derived from predicted probabilities are used to compute quantile thresholds. Both marginal coverage validity and class-wise (conditional) coverage were evaluated to ensure balanced reliability between stable and failure classifications under finite-sample conditions and minimal distributional assumptions. Interpretability is achieved through Shapley Additive Explanation (SHAP), Accumulated Local Effects (ALE), and Counterfactual Explanations (CFEs), which clarify the influence of geotechnical parameters and explain the reasoning behind each classification. The proposed AHA-RF-SCP framework establishes a transparent, statistically sound, and physically interpretable approach for slope stability assessment, promoting confidence and uncertainty-aware decision-making in geotechnical engineering practice.</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

Reliable slope stability prediction with guaranteed coverage: integrating conformal prediction and multi-method explainability

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月22日

查看全文
  • Shalom Adjei-Yeboah,
  • Mawuko Luke Yaw Ankah,
  • Bemah Ibrahim,
  • Yao Yevenyo Ziggah,
  • Constance Aning

摘要

The application of machine learning (ML) in slope stability assessment has been hindered by a lack of attention to uncertainty quantification and interpretability, which limits its adoption in engineering design. Prior studies have emphasized predictive accuracy without addressing the reliability and transparency necessary for decision-making in safety-critical contexts. This study introduces a novel framework for circular mode slope stability assessment that integrates split conformal prediction (SCP) with explainable ML to achieve statistically valid and interpretable predictions. The framework employs a Random Forest model optimized with the Artificial Hummingbird Algorithm (AHA) and implements SCP to generate calibrated prediction sets with guaranteed confidence. The dataset is divided into training, calibration, and testing subsets, where nonconformity scores derived from predicted probabilities are used to compute quantile thresholds. Both marginal coverage validity and class-wise (conditional) coverage were evaluated to ensure balanced reliability between stable and failure classifications under finite-sample conditions and minimal distributional assumptions. Interpretability is achieved through Shapley Additive Explanation (SHAP), Accumulated Local Effects (ALE), and Counterfactual Explanations (CFEs), which clarify the influence of geotechnical parameters and explain the reasoning behind each classification. The proposed AHA-RF-SCP framework establishes a transparent, statistically sound, and physically interpretable approach for slope stability assessment, promoting confidence and uncertainty-aware decision-making in geotechnical engineering practice.