In sandstone reservoirs with prolonged water injection and development, preferential flow paths tend to emerge during periods of high water content. This phenomenon significantly hinders the effective circulation of injected water, thereby compromising the efficacy of reservoir water injection and development. According to the causes and characteristics of preferential flow paths and the field data, permeability, effective thickness of reservoir, porosity, cumulative water injection, cumulative liquid production and water content are selected as the indicators of preferential flow paths, and then the improved CRITIC algorithm is used to weight the indicator samples, and a new machine learning algorithm XGBoost is introduced to train the calculation of the samples, and the CRITIC algorithm is finally established to identify the preferential flow paths in reservoirs. Finally, the CRITIC-XGBoost model for the identification of reservoir preferential flow paths is established. Examples of reservoir preferential flow paths monitored by inter-well tracers were organized, and the model was applied to identify reservoir preferential flow paths on the collected data, and the identification accuracy was calculated. The analysis comparing CRITIC-XGBoost with XGBoost, Random Forest (RF), and Support Vector Machine (SVM) revealed superior convergence and accuracy of the CRITIC-XGBoost model. This finding offers a novel and dependable approach for identifying preferential flow paths.