Research On The Recommendation System Of Combining User's Overall Preference And Local Preference

With the rapid development of the Internet,the phenomenon of information overload is becoming more and more serious.In this case,both the producer and consumer of the information will be affected to varying degrees.On the one hand,it is difficult for producers to information to ensure their product or information can be noticed by users;on the other hand,it is difficult for consumer to find their true favorite goals form vast amounts of information for products.To solve this problem,the recommendation system came into being.Traditional collaborative filtering technology used only the user's rating matrices on items to make recommendation.Because of the rating matrices were too sparse and the traditional way does not take ful y advantage of the many other features of users and objects,which led to a severe drop in recommendation accuracy for recommendat io n systems.In view of this situation,this paper makes full use of the rating data,puts forward two models to calculate the user's behavior preference,calculate the user's overall preference and local preference for the project respectively,and constructs a new hybrid recommendation model by introducing the control factor,and makes the project recommendation to the user.First,in view of the existing recommendation system based on implicit feedback,by the user as a result of interaction project build negative sample noise model is introduced into training situation,put forward a new way of dividing the positive and negative samples,by user rating of the project,project is counted as samples to high marks,low project for negative samples,by calculating the user on the positive and negative sample set for a project with the user's preference values of the ratio of the sum of all project preference value,get the user to a project's overall appetite.Secondly,in view of the existing recommendation system failed to dig deeper into the user preference for a project every characteristic of distribution,based on attention mechanism dynamic user interaction to its history learning list every characteristic of attention al ocation for each project,get the user on a particular project,the characterist ics of each order of preference for characterization of a new project,and characterization of this new project as the behavior of user preferences,ultimately through positive and negative samples are calculated respectively set on the user's preferences,distribut io n,average sum after local preferences of the customers.Finally,the user's global preference and local preference are combined through a control factor to obtain a hybrid model that can dynamically adjust local preference and global preference,namely Att+IFGP model.In addition,the performance of the model was tested.Through the comparison experiment,it was found that the Att+IFGP model output list was significantly improved in both HR value and NDCG value.