Research On Personalized Recommendation Method In Multi Heterogeneous Environment

With the rapid development of web2.0 technology,the Internet and e-commerce has been used widely.While enjoying more and more information service,people are confronted with the problem of information overload.Mass information makes it impossible for people to quickly find the information they want,and e-commerce personalized recommendation system can help users quickly find what they want or are interested in.As the most successful algorithm,collaborative filtering algorithm is the recommendation algorithm used most widely in recommendation system.It depends on the score given to the items by users.There is not enough score information of new users and new items that enter the system not long ago,thus the collaborative filtering algorithm cannot recommend new items and to new users.Apart from using the score by users,the thesis adds the rich social network information,the tag information and the time information to the recommendation system,and proposes the Bayesian-network-based collaborative filtering recommendation algorithm and the substance diffusion recommendation algorithm based on tripartite graph with tag weight.The main work of the thesis includes:(1)The thesis proposes the Bayesian-network-based collaborative filtering recommendation algorithm.The algorithm is based on friend network partitioning and Bayesian network.The algorithm first partitions the friend network,and then according to the tag information of users and the network partitioning result recommends to the user the tag which he or she is interested in(the tag represents the user's interest characteristic),and finally based on the Bayesian network gets the user's degree of interest in the items,thus the recommendation results.The algorithm is not only suitable for non-cold start users,but also can solve the problem of recommendation to the cold start users.(2)Given that the user's interest characteristic changes with time,in the thesis the tag is weighted according to the time when users tag.Because the user's recent tags are more convincing in revealing the interests of the user than the user's earlier tag,first,the thesis,on the basis of weighting function fitted based on the Ebbinghaus forgetting curve,obtains an attenuation function of the tag time,thus the weight of the tag marked at a time by the user can be obtained,then sets the initial resource in the user-tag-item tripartite graph network with the weight sum of tags marked at different times and places the initial resource on the tag,after that the substance is diffused in the tag-item bipartite graph.The algorithm is simple in principle and easy to implement.(3)Finally the thesis uses the Last.fm data set to compare the Bayesian-network-based collaborative filtering algorithm with user-based collaborative filtering algorithm in terms of precision,diversity and novelty.The result of the experiment shows that the Bayesian-network-based collaborative filtering algorithm is better at recommendation.The effectiveness of the substance diffusion algorithm based on tripartite graph with tag weight is verified in the Last.fm data set and sparser data set.The result of the experiment shows that compared with the tripartite graph with no weight sum,the algorithm can effectively improve the novelty of recommendation.Through the experiment it is also found that in the sparser data set the proposed substance diffusion algorithm based on tripartite graph with tag weight is better at recommendation than bipartite graph.