Abstract:The classification of printing oil is an important part of the document examination, in order to research the method of non-destructive and efficient identification of photosensitive printing oil. Taking the original spectral data of 33 different brands of photosensitive printing oil as the control group, after the original data were reduced by t-SNE and UMAP, three classification algorithms, XGBoost, SVM and MLP, were selected to determine the test set and training set in the ratio of 1: 4, and the original data and the reduced data were classified. At the same time, grid search and quintuple cross validation were used to optimize the performance and generalization ability of the model. The results show that the average accuracy of the above three classification algorithms for reduced-dimensional spectral data is higher than that of the original spectral data, and the UMAP-MLP classification model has the highest accuracy of 98%. The proposed classification model can be used to quickly distinguish the types of photosensitive printing oil.