Research On Recommendation System Based On Modeling User’s Long-term Preference

In today’s world,a mass of information is flooding everyone’s life.The advent of the information age not only improves the speed of social production and human life,but also produces a large amount of information,thus aggravating the problem of "information overload".For the majority of users,one of the main problems caused by information overload is that the huge amount of redundant information will greatly interfere with the accurate selection of information needed by users.How to get more useful information quickly from such a huge amount of data is a major problem existing at present.Recommendation system technology emerges as The Times requires,which can predict users’ needs and recommend them the content they are most likely to like,effectively reducing the visit time of people looking for interesting products or information on the Internet.There are many kinds of research on capturing user preferences（points of interest）in recommendation systems.In traditional recommendation systems based on the graph neural network,the method of capturing user preferences is to take the classical graph neural network model or its deformation as the way to learn the representation of nodes in the graph.Its basic principle is to input the bipartite graph into graph neural network to carry out message transmission between adjacent nodes,aggregate node neighborhood information,and capture deep node information in useritem interaction graph to learn user preferences and node characteristics of the item.However,existing works have no full consideration of the heterogeneity and dynamics in dynamic user-item interaction graphs or have no full application of both heterogeneous and dynamic information to the learning of users’ long-term preferences.Based on the above consideration,we propose a model of learning heterogeneous information and dynamic information from user behavior data based on the graph neural network.The main work of this paper is as follows:（1）Based on user interaction with the project information of heterogeneous information capture the problem of inadequate,designed for heterogeneous network figure HGCN neural model,the information transfer mechanism of graph neural model is used to mine potential heterogeneous information in the bipartite graph and help user node and project node to aggregate other types of node information around them.Based on the node features obtained,the graph attention network is used to learn the deeper heterogeneous information in the bipartite graph.By the above method,the heterogeneous information can be combined with the higher-order neighborhood information of nodes in the bipartite graph.（2）To capture users’ long-term preferences,this paper designs a graph neural network（Long-term Preferences of Users,LPre）for the dynamic graph of userproject interaction.In this model,multi-layer HGCN and graph attention network are used to capture heterogeneous information,and the self-attention mechanism is combined to capture dynamic changes of temporal information graph.Then we optimize the LPre model and propose LPre+ based on the consideration of heterogeneous information.Meanwhile,we learn the long-term preference characteristics of users and the long-term evolution characteristics of project nodes.In this paper,the LPre and LPre+ models proposed in this paper are tested for user-item interaction prediction on three user behavior data sets and compared with several mainstream baseline methods,the results show that the model presented in this paper has significant advantages in different data sets.