Improved Collaborative Filtering Recommendation Algorithm Based On Social Trust And Matrix Factorization

In today's society,the rapid development of Internet and Web technology has led to the rapid development of online video and audio systems and e-commerce systems,which greatly facilitates and enriches people's lives.But what followed was the exponential growth of the Internet data,vast amounts of data to the user and the enterprise has brought unprecedented opportunities and challenges,in the network of information overload lead to users cannot quickly find useful information on their own,directly led to the use of the information efficiency is greatly reduced,and how to find the required information in the shortest time is the user's most concerned problem.In this context,the recommendation system emerges with the trend of development and largely solves people's problems.The recommendation system can analyze and model users' previous information needs and interests based on their historical behavior data,and predict users' future interests and possible behaviors through the modeling results.Then the user is interested in the information and product recommended to the user or the information recommended to the user interested in it.Nowadays,recommendation system has been successfully applied in e-commerce,online video and audio fields.Recommendation systems have had an impact on every aspect of our lives.When we shop on Taobao,order takeout,listen to music and browse Tik Tok,recommendation systems have virtually recommended the content we are interested in.Collaborative filtering technology is one of the earliest technologies applied in recommendation system.Up to now,it is also one of the most successful and widely applied technologies in recommendation system.It can complete recommendation only by relying on user's historical behavior data,namely user-item score matrix,and has been successfully applied to various recommendation scenarios.There are many kinds of collaborative filtering techniques,among which neighborhood-based model and matrix factorization in cryptic model are the two most successful methods.In recommendation scenarios,although collaborative filtering performs well,its performance is usually limited to a certain extent due to data sparsity and cold start problems.Because the number of users and items in the network is huge,the number of items overrated by users and the number of users who overrated items account for only a small part,so the user-item rating matrix data is usually sparse.In addition,as new items are added or registered by new users from time to time in the network,the newly added items are almost not evaluated by users,and the newly registered users almost do not evaluate the items,so it is difficult to recommend new items to new users or to recommend new items to users who are interested in them.In order to alleviate these problems,scholars have proposed many different solutions one after another.Among them,the introduction of social trust mechanism makes user-item data well compensated,so it has become the most commonly used solution.However,because the number of users is huge and the chance of direct interaction between users is very small,the user trust network is often sparse.Therefore,on the basis of previous studies,this thesis proposes an improved collaborative filtering recommendation algorithm based on social trust and matrix factorization.The main work is as follows:(1)The implicit trust relationship between users is calculated based on the similarity between items graded by users,and a new trust network is constructed by combining the existing explicit trust relationship between users.(2)Calculation of trust matrix.In this thesis,the calculation of trust includes direct trust between users and indirect trust due to trust transitivity,and the final trust matrix is obtained by integrating the two.(3)Fuse the trust matrix into TrustSVD model to realize the final recommendation.In this thesis,a comparative experiment is conducted on two real public datasets of different sizes,Film Trust and Ciao,respectively,for ordinary users and cold start users.Experimental results show that the proposed algorithm reduces root mean square error(RMSE)and mean absolute error(MAE),which shows that the sparsity and cold start problems of the proposed algorithm can be effectively improved,and the theoretical results of existing recommendation algorithms are enriched to some extent.