Abstract: The vertical bearing capacity of pile foundations is closely related to the shear behavior at the pile–soil interface. The shear stress–displacement (t–z) curve is commonly used to characterize this behavior. The shear response at the pile–soil interface is influenced by numerous factors, and conventional t-z curve models often involve a limited number of variables, which restricts their ability to accurately represent the behavior. In recent years, machine learning has offered an effective solution for addressing multi-parameter and nonlinear geotechnical engineering problems. Building on existing physical insights, this study employs dimensional analysis to identify relevant parameters in dimensionless form. Using a dataset of t–z curves from previous studies, a hybrid model combining extreme gradient boosting (XGboost) and Gaussian process regression (GPR) is trained. The results demonstrate that the proposed t–z curve prediction model achieves high predictive accuracy. Furthermore, the computed load–displacement (F–z) curves of pile foundations based on the model show good agreement with experimental results, indicating strong generalization capability and practical applicability. Finally, using the SHAP method and parameter sensitivity analysis, the influence of various parameters on the model’s predictive performance is evaluated. It is found that dimensionless quantities related to normal stiffness, initial lateral pressure, and median particle size have significant impacts on the results, effectively capturing the influence of normal stress variation on shear stress during the shearing process.