INTRODUCTION: Traditional Chinese medicine (TCM) prescriptions are generally formulated by experienced TCM researchers based on their expertise and data statistical methods. METHODS: In order to predict TCM formulas for diabetes more accurately, this paper proposes a novel multi-source ensemble prediction method that combines machine learning ensemble techniques and multi-source data. In this method, the multi-source data contain datasets based on the components and targets (DPP-4 and GLP-1). Gradient boosting decision tree (GBDT), flexible neural tree (FNT), and Light Gradient Boosting Machine (LightGBM) algorithms are trained using these two types of datasets, respectively. The compound dataset from the TCMSP database is then used as testing data to predict and screen the active ingredients. The frequencies of occurrences of medicinal herbs corresponding to these three algorithms are obtained, each containing an active ingredient list. Finally, the frequencies of occurrences of the medicinal herbs obtained from the three algorithms using the component and target datasets are integrated to select duplicate drugs as the candidate drugs for diabetes treatment. RESULTS: The identification results reveal that theproposed ensemble method has higher accuracy than GBDT, FNT, and LightGBM. The medicinal herbs predicted include DISCUSSIONS: The results of network pharmacology show that myrrha can play a role in treating diabetes through multiple targets and pathways.