Abstract: A simple, fast and non-destructive classification method was established for leather shoe sole samples. X-ray fluorescence spectroscopy (XRF) was utilized to detect 55 leather boot sole samples of different brands, and seven elements (such as Cl, Ca, Zn, etc.) with high detection rates were selected as the classification indicators, and the samples were classified into four categories by combining with HDBSCAN clustering algorithm. It was found that the compositional features of the samples were significantly correlated with the brand positioning (e.g. sports, fast-fashion, etc.) and the production batch. Five models, including multilayer perceptron (MLP), random forest (RF), gradient boosted tree (GBDT), extreme gradient boosted tree (XGBoost) and support vector machine (SVM), were further used for classification validation, and the accuracy rates were 72.7% (MLP), 47.1% (RF), 64.7% (GBDT), 70.6% (XGBoost), and 82.4% (SVM), respectively. To improve the classification accuracy, the Stacking integration model was constructed using SVM, XGBoost and MLP as the base models and logistic regression as the metamodel, and the classification accuracy was improved to 94.1%. The results show that the integrated learning method can effectively extract XRF data features and significantly improve the classification accuracy of leather shoe sole samples.