Application And Research Of Collaborative Filtering Models In Recommender System

With the rapid development of Internet technologies and application services,the explosive growth of information has become an important feature of the era.However,people enjoy the welfare of information growth,while they have been facing with the problem of information overload.In this situation,how to get some interesting information for users from the sea of information rapidly and accurately has been a key problem highly concerned by many experts,users and service providers,at the same time,recommender system comes into being as an effective solution to the problem.As one of the most successful and widely applied recommendation technologies in recommender systems,the collaborative filtering model is the research object in this dissertation,and the aim of the dissertation is to solve the key problems in the recommendation process of collaborative filtering,for instance,the problem of data sparsity,cold-start problem,a trust measurement model,as well as a weight measurement model of a user and that of an item,in order to improve the recommendation quality of recommender system.Next,the dissertation will mainly describe the application and theoretical studies of collaborative filtering models in recommender systems from the following aspects:(1)It provides a comprehensive overview on the research background and the state of art of recommender systems in China and abroad,and analyzes some traditional and new problems existing in current systems.However,we want to put forward some new methods and theories to provide several effective solutions to these problems.And then,it describes the recommender systems in different application areas and their definitions,and introduces the composition modules and recommendation process of the system.Finally,it summarizes some commonly used recommendation models,including their working principles and scope of application,their advantages and disadvantages,as well as some performance evaluation metrics in recommender systems.(2)In view of the problem of data sparsity,it starts from the trust attributes and prefilles the rating matrix by introducing the conception of user credibility and establishing trust model.Secondly it coordinates the similarity between items from the perspective of ratings and user attributes by a self-adaptive balance factor,and proposes a collaborative filtering recommendation model based on trust model filling to get rating predictions of unrated items and final recommendations.Under the condition of data sparsity,the recommendation model can provide better prediction accuracy of ratings by enhancing the data storage density in rating matrix,which also shows that the proposed model can efficiently solve the problem of data sparsity.(3)Traditional collaborative filtering models establish similarity models only from the perspective of users or items,for which the system may face the performance degradation and suffer from the problem of data sparsity and the cold start-up problem etc.,due to the only source of information.In view of these drawbacks,it considers both users and items and proposes a collaborative filtering recommendation model combining users with items.Firstly the model optimizes the user-based and item-based similarity models,and then it respectively obtains the user-based and item-based prediction results with the category credibility of users and items.Finally,it ends the whole recommendation process by using a self-adaptive balance factor to coordinate both of the prediction results.(4)The importance is different for different users,and the same goes for items.However,traditional collaborative filtering models haven't given full consideration to the problem,but have given equal treatment to them in the recommendation process,which limits the system performance to some extent.In response to the drawback,it quantifies the weight of different users and that of different items,as well as the weight of an item for a user,and proposes four collaborative filtering recommendation models considering the weight of user and item.Finally the models,established on the basis of the weight rating matrix,show better performance in different data sets and different evaluation metrics,in particular with a reasonable cost in time.(5)There is a lot of contextual information in the rating matrix of the system,but the traditional collaborative filtering models haven't taken into consideration the information in measurement process of the similarity between users,but just perform some general operations on the ratings themselves.For the problem,it starts from the contextual information of ratings and carries out the statistical analysis of user ratings by rating singularity model.And then it combines the multi-channel diffusion model with the user similarity model,and proposes a collaborative filtering model fusing singularity and diffusion process and its extended model.Finally,the models demonstrate better performance and adaptability in the different data sets and different evaluation metrics,in particular with a reasonable time cost.(6)The multi-interest requirement of users is less considered in the recommendation process of systems,so it proposes a collaborative filtering recommendation model with item classification.The model firstly classify the rating matrix according.to the item categories,and then fills the matrix by an optimized similarity model iteratively to form a new user-item rating matrix with higher density of data.On this basis,we finish the global rating predictions and obtain item recommendations on the basis of the self-adaptive balance factor and local rating predictions.Finally the model not only demonstrates the excellent performance in the prediction accuracy and the category coverage of items,but also shows the bel:ter performance while the data sets become sparser.