Reservoir stimulation of shale oil is the key to subsequent development. However, there are numerous factors affecting reservoir productivity, and the main controlling factors of productivity for different horizontal wells remain unclear, leading to low accuracy in productivity prediction. To identify the main controlling factors of initial productivity of shale oil horizontal wells, taking the Chang 7 Member of the Ordos Basin as the research object, Pearson correlation analysis, principal component analysis, and grey relational analysis were used to screen and reduce the dimensionality of productivity-influencing factors. Based on the screening results, a machine learning prediction model was established to reasonably predict the initial productivity. The results show these as follows. By weighing the correlation degree between various factors obtained by different methods, the number of fracturing stages, proppant dosage per meter, oil saturation, soaking time, pumping depth, and porosity are finally determined as the main controlling factors of productivity in this area; taking the screened main controlling factors as input parameters, the integrated model established based on 5 single machine learning algorithms has high prediction accuracy, with R², RMSE (root mean squared error), and MRE (mean relative error) reaching 0.958 1, 0.173 8, and 1.119 1% respectively; compared with single machine learning algorithm models, the integrated model can better predict the initial productivity of horizontal wells and improve prediction accuracy. In the future, the model algorithm can be optimized by combining multi-source data to enhance the prediction effect, which has important guiding significance for the optimization of actual shale oil horizontal well development.