AbstractPrecast segmental concrete beams (PSCBs) are being increasingly applied in bridges worldwide benefitting from the advantages of accelerated bridge construction. It is of importance to accurately predict the direct shear strength (DSS) of precast concrete joints (PCJs) for ensuring the safe structural design of PSCBs. However, existing prediction models of PCJs’ DSS are deemed inaccurate and unreliable when numerous parameters are varied in wide ranges. This study aims to establish an accurate and reliable prediction model for PCJs’ DSS using a machine learning algorithm called support vector regression (SVR). A PCJs’ DSS database of 304 test results with 23 input parameters was assembled from the literature. A model training procedure was conducted through stratified train-test split, feature scaling, feature selection, and two-step grid-search hyperparameter tuning. A new correlation matrix–based feature selection method was proposed, and three SVR models with different feature combinations were trained for validating the selection method. The trained SVR models were experimentally validated and compared with six existing mechanical models through two groups of performance indicators. A reasonable interpretation for the SVR model with the selected features in the proposed selection method was done using the combination of partial dependence (PD) and individual conditional expectation (ICE) plots. The results show that the SVR algorithm can be deemed feasible to accurately and reliably predict the DSS of PCJs. The proposed feature selection method is beneficial to the prediction performance of the SVR model. It is impossible for the typical mechanical models to achieve a similar prediction performance of the SVR model. The influence of each input parameter on the DSS of PCJs is recognized and depicted, which can offer useful information for further developing new mechanical models for predicting the DSS of PCJs with higher prediction performance in future research.