Research On Recommendation Algorithms Against Cold-Start Problem

With the rapid development of Internet technology,the information is exploding.How to get the information needed from overloaded information has become an urgent problem,and the recommendation system is a representative solution to the problem.The recommendation system connects users and information by helping users discover the information actually needed or showing the interested information for the users,so as to realize the win-win between the information consumers and the information producers.The recommendation system involves many algorithms,and the collaborative filtering recommendation algorithm is the classic algorithm.However,the recommendation algorithms are all based on a great deal of data about users and items,while the size of data could not meet the requirements,the data sparsity will be occurred,which can directly leads to the cold-start problems of recommendation system.The cold-start problem will make the recommended accuracy of the recommended algorithm drop considerably,and has a direct impact on the user's experience and trust of recommended system.Nowadays,the cold-start problem has become the research focus in the field of recommendation technology.In this thesis,some researches of cold-start problems caused by data sparsity are done from two aspects: detecting cold-start problem and solving cold-start problem.To the detection of cold-start problems,a "scene switch" method based on state detection is designed,which can be used to detect status of users and items in order to select the appropriate processing algorithm based on the detected state.And it is proved by the experiment that the method has certain detection accuracy and preprocessing effect for cold-start problem.In order to solve the cold-start problems,the improved methods are designed for collaborative filtering recommendation system and contextual recommendation system respectively,they are Collaborative Filtering Algorithm Based on User Similarity and Rating Information(SR-CF),and Context Recommendation Algorithm Based on Factorization Machines(FMs).The SR-CF method improves the similarity of users' rating to a hybrid similarity which combines the similarity of users' basic information,similarity of seed-sets rating and similarity of users' rating with using weighted summation.The FMs method is based polynomial regression,which uses the factorization machine to factorizing the interaction between features to reduce the complexity of the algorithm.The experiments are designed and the experimental results show that the methods not only improve the recommendation accuracy but also the time efficiency.In summary,the algorithms designed in this thesis can not only alleviate the negative effect of cold-start problem,but also increase the accuracy and efficiency.The work done has certain value both on theory and practice.