Abstract: The support services related to reference books is one important task of university libraries. The prerequisite of carrying out that is to identify the frequently utilized reference books. Based on machine learning methods and massive dynamic data about user behaviors, this paper first constructs seven-dimensional feature sets about borrower types, borrowing time and utilization rate to establish the identification model of frequently utilized reference books. In the experiment session, Support Vector Machine, Decision Tree, Random Forest and XGBoost algorithm are selected for comparison. Among them, XGBoost algorithm yields the best experimental performance. To be specific, its precision, recall and F1-score are 0.849, 0.906, 0.876. In summary, we obtain good identification result in this paper.