Abstract: The existing personalized recommendations often ignore the effect on the users’ behavior caused by the time information, resulting in a lower forecast accuracy. According to the user attribute information and user rating information, this paper sets up a mixture recommendation model (TLDA-CF) based on the temporal context-aware LDA model and collaborative filtering model. It improves the recommendation efficiency through offline and online recommendation; respectively establishes the LDA models based on the user rating information and the project’s access amount in every period of time to solve the problem of data sparsity; sets a dynamic weight to balance the user selection trend to improve the accuracy; and uses user clustering and collaborative filtering algorithm to predict the users’ preferences for those new users with no rating information, so as to solve the cold start problem. The results of the proposed algorithm’s experiment on MovieLens datasets indicate that this algorithm is superior to the traditional LDA model on the recommendation recall rate, accuracy, and F1 values.

Key words: personalized recommendation, collaborative filtering, LDA, recall, precision