Abstract: In order to solve the problem of imbalanced classification and high similarity of logging responses among different rock types in tight sandstone reservoirs, which seriously affects the automatic recognition effect of machine learning rock types, this paper proposes the SSMO-SSA-LGBM model: combining SVM algorithm with oversampling technique SMOTE to form the SSMO model, balancing the samples with fewer lithology data in the training set to obtain new composite samples, and combining them with the original training set to form a new training set for training and constructing the LGBM model. Due to the use of many hyperparameters during LGBM model training, the sparrow optimization search algorithm SSA is used to optimize its hyperparameters and seek the best parameter combination, forming the SSMO-SSA-LGBM combination model. Based on the lithological data of Yan10 tight sandstone in Huachi S area, an SSMO-SSA-LGBM model was trained and constructed, and compared using models such as KNN, Adaboost, and random forest. Research has shown that after oversampling by the SSMO model, the LGBM model enhances its performance in minority class recognition; The SSA algorithm global optimization search obtains the optimal hyperparameter of LGBM through fewer iterations, achieving optimal predictive performance. After multiple runs on the test and validation sets, the accuracy of the prediction results remains stable at over 94%, verifying the good application effect of the model in Huachi S area.