Research On Collaborative Filtering Recommendation Leveraging Social Information

With the popularity of mobile terminals,mobile Internet application has become an indispensable part in people's daily life.People can easily access the Internet,use Internet services,and publish their own content.The number of social network's users,the amount of content and services are also growing steadily.People on the social media network(such as micro-blog,WeChat,Zhihu)produce and consume content constantly.More and more people are getting news from the information streams based on social media networks.Zhihu is an online Q&A community,where the user's news feed displays on the website home page.The information stream will become increasingly intertwined and complicated with the growing connections among users.For those who followed many users,they can get thousands of messages in a very fast manner every day.So personalized recommendation systems are used by major platforms to filter contents that could interest specific users.The content of Zhihu is made up by a number of questions.Each question can be tagged with different topics,and each question comes with an answer list on the question page.Users can add their own answers,or vote and comment an answer from other users,follow the question or in some other ways to express their attitude towards the answer.The content of Zhihu is propagated through the relationships between users.In this thesis,we propose a graph recommendation model,based on random walk,which brings users,topics and questions together,to build users' interests graph.Additionally,a type of time nodes are added into this graph to take users' long-term and short-term interests into consideration.Moreover,the user's social relationship is also integrated into the algorithm to identify users' potential relations with the help of Strength of Weak Ties Theory.Finally,we would be able to create a personalized recommendation list based on the analysis of user's social relationships and their feedbacks.In Zhihu,there are users who are willing to share their knowledge and experience,and they are content producers,enriching the platform with what they have created.In the meantime,there is another group of users we call content consumers.After reading the readily available contents,some of them would share their own point of views by voting certain answers or forwarding the feed,which are recorded by the platform as an archive of user behavior.Moreover,there is another group of users who read more and seldom make any answers.Except from their social relationships,no other behaviors could be collected.In this thesis,we propose a matrix factorization model based on the text content that is designed to make recommendations to the users described above.The relationship between users and questions is described by the "user-topic" matrix.Matrix factorization is used to obtain the "user-topic" feature matrix,and then a large number of questions are mapped to a small number of topics through the relationship between topics and questions.At the same time,for users who are less active,we build the user topic model by adding the content of the answers given by people they have followed,in this way,we could enrich the dataset.In addition,considering the real situation,we have designed a TopN recommendation method,which utilizes a threshold.This method can flexibly adjust the length of the user's recommendation list,and recommend the most suitable content to users in different scenarios.This method has been proven more effective with the real dataset collected from Zhihu,and it works better than the baseline algorithm.