Calcium sulfoaluminate (CSA) cement mixture design is challenging due to the influence of multiple features on its unconfined compressive strength (UCS). Consequently, the relationships between input features and the UCS exhibit non-linear behavior, making it difficult to understand using experimental methods alone. Therefore, for the first time, this study constructed non-linear ensemble machine learning (ML) models on a dataset compiled from experimental literature to accurately predict the UCS of CSA cement mixtures. After applying feature selection techniques, four different ensemble models were built on the modified datasets to predict the UCS. The extreme gradient boosting model built on the dataset modified by the least absolute shrinkage and selection operator method achieved the best prediction accuracy (coefficient of determination; R2 = 0.95) on testing data. Finally, the SHapely Additive exPlanations analysis could interpret the selected ML model both quantitatively and qualitatively, by explaining the independent relationships between each input feature and UCS.