Research On Personalized Recommendation System Based On Context-aware And Feature Interaction Modeling

With the advent of the era of big data,the amount of information stored on the Internet is growing rapidly.People enjoy the benefits brought by big data,but also face the problem of being surrounded by data information,that is,the problem of "information overload".Personalized Recommendation System can spontaneously learn the user's behavior,understand the user's information,and learn the user's interest.Therefore,Personalized Recommendation System has become one of the most effective ways to solve the problem of "information overload".Context-aware Recommendation System is a typical personalized recommendation method.The advantage of Context-aware Recommender System is that it can fully consider the influence of context environment on users' interests.Therefore,compared with traditional Recommender System methods,Context-aware Recommender System has higher recommendation efficiency,which has attracted widespread attention.Current Context-aware Recommendation algorithms simply map multi-field features(user,item and context information features)to the shared implicit semantic space,and then connect them to Deep Neural Networks or other specially designed networks.In these studies,multi-field features are combined simply and unstructured,so the interaction of modeling features is not flexible and explicit.In addition,mapping context and user(or item)to a shared implicit semantic space to capture the interaction between context and user(or item)can only learn the simple interaction relationship between context and user(or item),but the collaborative signal hidden in context-user(or item)interaction is not learned,so the recommendation efficiency is poor.In order to solve the above problems,we have done the following work:(1)Graph Neural Network and Context-aware Based User Behavior Prediction and Recommendation System ResearchWe propose to use graph structure to directly represent multi-field features,and construct context-user(or item)feature interaction graph.Each node in the graph represents a field feature,and different field features can connect and interact through the edges in the graph.Therefore,the task of modeling context-user(or item)interaction can be transformed into modeling the node interaction on the graph.Therefore,we design a new model——Context-aware Graph Neural Network(CA-GNN).CA-GNN uses graph structure to model the interaction between multi-field features,which can model the complex interaction between different field features in a flexible and explicit way.The experimental results show that the CA-GNN model is better than other benchmark methods in Food and Yelp datasets.Compared with the more advanced model before,CA-GNN model improves the RMSE index by 2.2% and the MAE index by 1.8% in the Food dataset,and CA-GNN model improves the RMSE index by 1.4% and the MAE index by 1.3% in the Yelp dataset.(2)Research on Context-aware Recommendation System Based on Attention Interaction Graph Neural NetworkWe propose to construct a bipartite graph structure from the context-user(or item)feature interaction relationship,and use Graph Convolution Network to optimize the embedding vector of user and item.At the same time,we also use the Attention Mechanism to learn the importance of different context features on users(or items).Therefore,we design a new model — — Attention Interaction Graph Convolution Network(Ai-GCN).Ai-GCN can model the high-order connectivity in context-user(or item)graph,and effectively inject the collaborative signal between context and user(item)into the embedding process in an explicit way.Experimental results show that our AiGCN model outperforms other benchmark models on Frappe dataset.Specifically,compared with the more advanced model before,the Precision@5 index increased by0.2%;the Precision@10 index increased by 1.2%;the Recall@5 index increased by2.4%;the Recall@10 index increased by 8.6%.The CA-GNN model and Ai-GCN model established in this paper both have the advantages of "pervasive computing" and "personalization".By considering the influence of context,they can further improve the recommendation accuracy and user satisfaction.In the future application scenarios,CA-GNN model can be well applied to the scoring prediction task of recommendation system,Ai-GCN model can be well applied to the personalized sorting task of Recommendation System.