Loess Collapsibility Prediction and Influencing Factor Analysis Using Multiple Machine Learning Algorithms in Xi’an Region

Collapsibility is a fundamental geotechnical property of loess that critically affects its engineering behavior. In this study, a comprehensive dataset comprising 9041 experimental records on the physical properties and collapsibility of loess from the Xi’an region was compiled. Six parameters were selected as model inputs: sampling depth (H), water content (w), plastic limit (wP), plasticity index (IP), compression coefficient (a1–2), and compression modulus (Es). Based on these inputs, prediction models for the loess collapsibility coefficient (δs) were developed using Gaussian Process Regression (GPR), Gradient Boosting Machine (GBM), Support Vector Regression (SVR), Radial Basis Function Neural Network (RBFNN), Classification and Regression Tree (CART), and Feature Tokenizer Transformer (FT-Transformer). Among these, GPR demonstrated the best predictive performance, achieving the lowest error (RMSE = 9.88 × 10−3) and the highest accuracy (R2 = 0.844). Additionally, the coverage proportion of the 95% confidence interval of the GPR predictions reached 0.949. SHapley Additive exPlanations (SHAP) analysis for GPR further revealed that the compression coefficient exerted the greatest influence on δs (0.0149), followed by compression modulus (0.0080), water content (0.0068), plasticity index (0.0061), sampling depth (0.0061), and plastic limit (0.0052). The GPR-based prediction model offers significantly higher predictive accuracy than empirical models. The developed models provide a robust technical framework for the rapid estimation of loess collapsibility in the Xi’an region.