Research And Application Of Personalized Recommendation Algorithm Based On Real-Time User Interest Preference Modeling

With the explosive growth of the amount of data on the Internet,it is difficult for people to find useful information in the massive amount of information quickly,which leads to the emergence of information overload.The emergence of the recommendation system is considered to be the most promising effective means to solve the information overload.It can predict the next behavior for users by mining the historical preference information on users.Among them,it is necessary to model user behaviors in order to portray user interests accurately in order to better understand users.However,the existing user interest modeling methods models analysis rely too much on the use of rating data.Due to the sparseness of user rating set,this modeling method cannot solve the cold start problem in the recommendation system well,and the user's interest tend to drift significantly over time.This master's thesis uses deep learning,density peak clustering,collaborative filtering and other technologies to study the recommendation algorithm based on the user interest model,and solve the problems of cold start and user interest drift in the recommendation algorithm.The research content mainly includes the three aspects as follows:(1)The user's interactive record information is used in the traditional user modeling method,which cannot solve the cold start problem that is often encountered.In order to solve this problem,the attribute information provided by the user when registering in the new system is used in this master's thesis.The user attribute information is embedded and represented as a user attribute vector,using One-hot encoding.Afterwards,the Auto Encoder network is used to reduce the dimension and combine the user attribute vectors to obtain a user model that integrates rich attribute message.It is verified through experiments that the method in this master's thesis can obtain a richer user model and solve the problem of cold start.(2)The influence of the time factor on the user's interest and the popularity of the item is ignored by most traditional recommendation algorithms when learning the interactive information between the user and the item in a unified manner.In order to solve this problem,the time series information recorded by the user interaction is used in the design of the user model.Perform deep learning on time series information and use LSTM to obtain the user's current interest preference vector.Finally,it is used with the user model combined with rich attribute information to obtain the final user model.After that,clustering the obtained user model vectors with a clustering algorithm based on density peaks,and calculating the user similarity in the same cluster,can greatly reduce the time consumption of calculating the user similarity.The data set is divided in the order of item interaction time,and higher weight values are assigned to items that have recently interacted,so as to obtain a list of items most suitable for the target user's recent interests.It is verified through experiments that the recommendation quality of the recommendation algorithm designed in this master's thesis is better than that of the comparison algorithm,on the Movie Lens-100 K,Movie Lens-1M and Book-Crossing datasets.(3)Based on the recommendation algorithm proposed in this master's thesis,the application of recommendation system based on educational resources is realized.and verify the rationality of the proposed algorithm and better recommendation effects effectively.