The efficiency of traditional manual labeling of user profiles is relatively low, so using algorithmic models to predict labels has become a popular research direction in user profiles. However, current mainstream machine learning algorithms have not been able to deeply explore the complex relationships between features, and the prediction performance is still unsatisfactory in high-dimensional and sparse features. Hybrid algorithms can often combine the advantages of various algorithms, overcome shortcomings to a certain extent, and improve prediction accuracy. This article proposes a two-layer ensemble learning framework based on XGBoost algorithm. Based on the characteristics of user queries for word materials, this article analyzes and compares the advantages and disadvantages of commonly used document vector representation methods. In the model, XGBoost algorithm is used as a classifier to construct a user profile model. The experimental results show that this method can effectively reduce feature dimensions, enrich the diversity of base classifiers, and efficiently and accurately achieve the prediction of multi-dimensional attributes by the user profile model.