A Study Of Cross-domain Recommendation Algorithm Based On Knowledge Transfer

In the age of Internet,the accumulation of a large amount of information makes it difficult to quickly and accurately find those contents they are interested in.Recommendation system,to some extent,solves the problem of information overload,but it is difficult for traditional recommendation system to solve some problems,such as cold start and data sparseness.With the development of Internet,the information between different fields is shared and complementary,which brings the opportunity to solve the problems about cold start and cross-domain recommendation.In order to improve the accuracy and diversity of results from cross-domain recommendation as well as the utilization rate of cross-domain information resources,this paper proposed two cross-domain recommendation algorithms based on inter-domain knowledge transfer.The main works are as follows:(1)We firstly analysized these users between source domain and target domain,and proposed the cross-domain recommendation algorithm(User interest-based transfer,UIT)based on user's transfering interest in different scenarios in which overlapping users of domains exist.From the perspective of users,users' interest will be reflected in different fields and users' friends in different fields are maybe different.Based on this,through the users' friends in the information-rich fields we can transfer the users' interest to the target recommendation domain where relative information is sparse,we will fill the rating matrix in source domain,then calculate the similarity by matrix factorization,finally we will obtain the fusion algorithm which combines the users' interest in source domain and modified algorithm in target domain.(2)Considering that the UIT algorithm needs to meet the conditions with overlapping user and inter-domain user group has a cross section on the feature level,and we further proposed a cross-domain recommendation algorithm called SKP(Sharing knowledge pattern)based on the shared knowledge pattern.By analyzing the user-item-rating data in each domain,we can obtain users' potential characteristics and items'potential characteristics,and then on the base of these,through the use of clustering method we can extract knowledge pattern of the collection of users and items in each domain,and finally make full use of individual knowledge pattern and the sharing knowledge pattern between domains to provide final recommendation results.(3)In the Spark cluster environment,we implement parallelism algorithm proposed in this paper and related compared algorithm,as well as some optimization.The experimental results show that compared with collaborative filtering algorithm in single domain and the current cross-domain algorithm,the algorithm proposed in this paper has lower RMSE,higher accuracy,recall rate and F1 measure,at the same time,the algorithm in this paper is extensible and real-time under the verification of Spark clustering nodes.