Research On Trust Collaborative Recommendation Based On Recurrent Neural Network

In real life,there is a certain sequential relationship between users' behaviors,so the recommendation system should mine the dependency relationship between users and items according to the sequence of users' behaviors,so as to generate recommendation results.This recommendation method,which predicts the user's preference at the current moment based on the user's behavior at the previous moment,is generally called Sequential Recommendation.The existing Sequential Recommendation model generates recommendation results by modeling the sequential dependence of user-item interaction,but there are still some problems in existing models:First of all,the existing Sequential Recommendation models often fail to properly allocate the influence of each user-item interaction on the user's interest and preference.Instead,they only regard each user's behavior as a simple click,without considering that the user-item interaction behavior also reflects the user's interest and preference at a deep level.This problem leads to the lack of personalized user characteristics in recommendation model modeling and the generated recommendation results often can not satisfy users.Secondly,the current recommendation system is filled with all kinds of misleading information.Such misleading information not only deceives users and leads them to buy products that do not conform to their interests and preferences,but also causes the recommendation system to fail to learn accurate user expressions,reduce the accuracy of recommendations and affect users' trust in the results of recommendations.Given the problems existing in the current Sequential Recommendation Models,we conducted the corresponding research.The main research work and achievements are as follows:1.We propose a fusion of item-side information(the expression of item-side information is different in different scenarios: in the e-commerce scene,the item-side information includes:type,brand,price,etc.,and in the film scene,including the type of film,protagonist,music,etc.)of GRU Recommendation Model(FIGRURec).This Recommendation Model models the behavior sequence of users through the Gated Recurrent Unit(GRU)and abstracts the change rules of users' interests.In order to enhance the personalized representation of users,the attention mechanism is used to assign the corresponding weight to the items that users interact with.By combining user behavior sequence,item-side information and user characteristics,user modeling is carried out to ensure that the model can generate more personalized Top-K recommendations while learning more accurate user characteristics.2.Given the misleading information in the recommendation system,a component that can identify abnormal users is proposed.This component can identify abnormal users by comparing their actual ratings with their predicted ratings,and then update the characteristic representation of abnormal users to ensure the accuracy and robustness of the system.3.The performance of the FIGRURec model and the abnormal user identification component was verified on multiple data sets.The results show that our model can improve the performance of 6%-8% compared with other Sequential Recommendation Models,and the proposed abnormal user identification component can effectively improve the ability of the recommendation system to deal with misleading information.