To solve the problem about low accuracy of traditional information recommendation method, this paper introduces the user portrait as the recommended basis, and proposes the recommendation method based on dynamic user portraits after further studying about the text classification and user behavior. By dynamically analyzing the user's historical data and predicting the user's interest trends, this method achieves dynamic recommendations. The off-line experiment improves that this method has some advantages in predicting user preference changes compared with the traditional label-based information recommendation, and it improves the recommendation accuracy.