Modeling User Long-term And Short-term Preference For Sequential Recommendation

In a network platform with a large amount of information,the recommendation system can help users effectively take advantages of the information,quickly screen out their desired products,and improve the search efficiency.The core of the recommendation algorithm is to model user preference based on the user's historical interaction behavior and auxiliary information.Besides,the recommendation system provides users with a list which meets their expectations.In real life,users' interest and preference often changes with time.Therefore,sequential recommendation algorithm capable of capturing dynamic changes from the user's historical behavior has become a research hotspot.Most of the sequential recommendation algorithms are barely considering the user's recent interaction behavior without obtaining the intrinsic relationship between these behaviors.Some of the state-of-the-art algorithms just linearly model the user's recent behavior,which is weak to learn the true preference of user with a flexible order sequence.How to obtain the potential dependencies between users' historical behaviors and capture the users' changing preference over time is an important issue that needs to pay attention at present.In view of the above problems,the main contents of this paper are as follows:1.First,a joint model based on user long-term and short-term preference is proposed,which embeds the sequence of items clicked by the user into a deep neural network.LSTM model is used to receive interactive information and learn the potential characteristics of the neural network to generate the user state at each moment.Next,the self-attention machine is used to estimate the recent user behaviors for enhancing the weight of related items,and so that the prediction result of user replacement at the next moment will be output.This model can learn the complete product sequence,and simultaneously obtain the long-term interest and short-term preference of the user.2.In order to add the application of product information in the recommendation system,the correlation information within neighborhood products is calculated in the Yoochoose dataset through random variable analysis and hypothesis testing.The statistical results prove that the neighborhood products own a predictive effect on user preference at the new moment.After that,a product information collaboration model is proposed.Based on the user's historical behavior information,the dependency relationship between products is calculated without other auxiliary information.According to the related product items,a recommendation list will be generated to the user,and the recommendation result is integrated into the user's long-term and short-term preference model.The recommendation ability of the sequential model is enhanced from the perspective of product collaboration.3.By performing the recommendation task on the real datasets,this paper verifies the validity of the model and discusses the effect of the parameters set in the model on the experimental results.