An integrated SHAP-MCDM approach for slope stability prediction based on machine learning algorithms

In machine learning (ML)-based slope stability prediction studies, feature importance results often vary across different algorithms, leading to inconsistent interpretations. This issue arises because the importance of features differs depending on the algorithm applied within the same study. To address this challenge, this study proposes a novel methodology for obtaining a final, unified ranking of features by combining the feature importance rankings of various ML algorithms using a Multi-Criteria Decision-Making (MCDM) technique. This approach ensures a consistent and reliable feature ranking derived from the results of successful ML models. Furthermore, the study demonstrates how performance indicators of ML algorithms can be translated into criterion weights within the MCDM framework. Hyperparameter optimization was applied to the ML models, achieving accuracy rates between 70% and 92.5%. Successful algorithms were analyzed using SHapley Additive Explanations (SHAP) to evaluate feature importance, and the results were integrated into the proposed SHAP-MCDM methodology. The MULTIMOORA method, a well-established MCDM technique, was employed to combine the SHAP rankings. The results confirmed that a final feature ranking could be derived by merging different SHAP rankings of ML algorithms using the proposed SHAP-MULTIMOORA approach. The study also identified key features like cohesion, internal friction angle, and slope height, which significantly influence slope stability prediction. This methodology both contributes to the integration of SHAP rankings and advances prediction accuracy by calculating criterion weights of ML algorithms based on multiple performance metrics. The proposed approach has a broad applicability, improving both classification and regression-based prediction tasks in various domains beyond slope stability.