Abstract: The residual shear strength of rock joints is a critical parameter for assessing the long-term stability of rock masses. To address the issues of insufficient consideration of influencing factors and limited validation data in existing prediction models, this study systematically identified eight multi-dimensional key influencing factors through red sandstone joint shear tests and analysis of existing research. A multi-source dataset comprising 630 sets of test data was constructed by collecting joint test data of different lithologies. Using machine learning methods, three support vector regression (SVR) models and one deep learning (DL) model were developed to capture the nonlinear relationship between residual shear strength and the key influencing factors. Based on a dataset comprising joint shear test results from different lithologies, four machine learning models were compared with the Ban and Peng models. The results indicate that all machine learning models exhibit higher prediction accuracy than the Ban and Peng models, with the DL model achieving the best performance and a prediction error of only 8.09%. This demonstrates that the machine learning models possess strong predictive capability under varying lithological conditions. Model validation using the results of red sandstone shear tests further confirms the effectiveness of the machine learning prediction model.