Three-level evaluation method of cumulative slope deformation hybrid machine learning models and interpretability analysis

Abstract

The study aims to address the issues of limited evaluation metrics and low visualization for Machine Learning (ML) models in Cumulative Slope Deformation (CSD) prediction. Firstly, to establish a representative and balanced CSD dataset, the Grey Relational Analysis (GRA) and Mutual Information (MI) were used to determine the main controlling features influencing CSD. Secondly, combining four typical ML algorithms Support Vector Regression (SVR), Extreme Learning Machine (ELM), Long Short-Term Memory (LSTM), and Gated Recurrent Unit (GRU) with an Improved Particle Swarm Optimization (IPSO) algorithm to develop hybrid ML models. Moreover, a three-level evaluation method is proposed considering prediction accuracy and error, prediction uncertainty, and prediction robustness to determine the optimal hybrid ML model. Then, the effectiveness of the proposed method is validated using data from a diatomaceous earth slope on the site. Finally, the SHapley Additive exPlanations (SHAP) method was used to perform interpretability analysis on the optimal hybrid ML model in three steps: Moment-specific and global interpretation, Overall importance analysis and contribution analysis, and Dependency analysis. The results indicate that: 1) Based on the combined feature analysis algorithm of GRA and MI, the main controlling features affecting CSD were the three-day cumulative rainfall, daily rainfall, daily moisture content, daily deformation increment, and daily moisture content change. 2) The IPSO algorithm has a significant advantage in improving the prediction accuracy of the ML models. 3) The calculated Comprehensive Evaluation Index (CEI) values for each hybrid ML model are as follows: IPSO-GRU (1.980) < IPSO-SVR (2.380) < IPSO-LSTM (2.564) < IPSO-ELM (2.793), which indicated that the IPSO-GRU was the optimal hybrid ML model. 4) In the IPSO-GRU model, the top three features in order of overall importance are Previous CSD, three-day cumulative rainfall, and daily rainfall, and the feature most closely interacting with three-day cumulative rainfall and daily rainfall is the Previous CSD. The research findings can provide a reference method for selecting the optimal ML model and offer a precise, interpretable, and reliable predictive model for CSD.