In order to solve the data sparsity problem and improve the recommendation accuracy, we proposed a new matrix factorization XGBoost(MFXGB) recommendation algorithm which combined the matrix factorization method and the XGBoost (Extreme Gradient Boosting) algorithm. SVD++ algorithm (SVD, singular value decomposition) was used to fill the user-item score matrix to lessen the influence on the accuracy of the algorithm due to too many missing values and XGBoost was used to build a supervised learning model to predict the user's score. To reduce the computation time, feature extration based on the K-means clustering method was proposed to train XGBoot. The proposed MFXGB algorithm was applied to MovieLens dataset for experimental analysis and the results show that the recommendation accuracy was improved by 8.91%,10.18% and 11.79% respectively, compared with the three traditional algorithms.