Due to the characteristics of low permeability and productivity of tight oil reservoirs, hydraulic fracturing has been widely applied to improve oil recovery. According to different geological and fracturing parameter changes, predicting the recovery rate after modification is instructive for fracturing modification. At present, the theoretical model of tight oil recovery after fracturing affected by multiple factors is difficult to accurately predict the change of oil reservoir recovery after fracturing in real-time according to the fracturing method and parameters. To further improve the prediction accuracy of tight oil recovery prediction, machine learning is introduced to make predictions and has certain improvements have been made based on the eXtreme Gradient Boosting Algorithm (XGBoost) and the Support Vector Regression Algorithm (SVR) to obtain the variable weight combination model XGBoost-SVR. The combined model can complement both single model’s advantages to avoid the range error caused by a single model parameter, and thus increasing the model prediction error tolerance rate. Firstly, factors affecting the recovery of tight oil are collected and sorted, and the relevant original data sets have been established after analyzing the influence of geological factors, reservoir factors, and engineering factors on the recovery factor; secondly, the preprocessed data sets are inputted into the SVR single model and the XGBoost single model for training separately, and the single model prediction value is obtained; finally, an adaptive variable weight combination method based on residuals is used to establish the XGBoost-SVR combination model, which can obtain the final prediction results of each model, and clarify the factors affecting the recovery factor and the weight ratio of each factor. It is shown from the prediction results that compared with the SVR and XGBoost single models, the combined model has a prediction accuracy of 94.63%, which reflects better adaptability.