Research On Personalized Recommendation Techniques In Social Media

In the age of the Internet,with the development of information techniques,a variety of social media have emerged and grown rapidly.The arising of social media has changed people's social lives significantly.Existing social media not only meet people's need for social association,but also satisfy users' information demands by facilitating the creation,sharing,and spread of information.However,with the rapid development and widespread popularity of social media,the number of social users and social contents has increased dramatically,which leads to the explosive growth of social data.These massive social data have brought a troublesome information overload problem to social users,which means social users feel that it is more difficult to find interesting contents from huge volumes of social data.To solve the information overload problem,personalized recommender systems,which recommend contents of interest to users according to their preferences and help users find valuable information,have been widely employed in different kinds of social media.Due to its high application value and significant commercial value,personalized recommendation in social media has received extensive attentions from both industry and academia.The various types of social media and the diversification of social data give new opportunities and challenges to the personalized recommender systems in social media,although researchers have already made some achievements in this field,there is large room for improvement in many aspects,such as the recommendation accuracy,the utilization of the social media's characteristics,and the exploiting of the social contents.Therefore,in order to exploit the own characteristics of personalized recommendation in social media and overcome the shortcomings of existing works,we make in-depth researches on three different personalized recommendation problems in social media:personalized recommendation based on items'multiple attributes,personalized recommendation incorporating user relationships,and personalized social relation recommendation exploiting the heterogeneous network.The major contributions of our researches are listed as follows:(1)For the problem of personalized recommendation based on items'multiple attributes,we make a study on the personalized new social event recommendation in event-based social networks based on a social event's multiple attributes and propose a unified hybrid model which integrates user interest,organizer influence,and geographical preference.In the model,based on all the contents of a user's attended social events,we employ a topic model to obtain the user's interest.Meanwhile,we construct a user-organizer matrix to represent all the organizer influences on users and propose a popularity-aware probabilistic matrix factorization method to infer the missing values in the matrix.In addition,we propose a probabilistic model which considers both the locations of a user's attended social events and the numbers of events the user has attended at different locations to model the user's geographical preference.The experimental results on real datasets demonstrate that our proposed recommendation model is effective in recommending new social events in event-based social networks.(2)For the problem of personalized recommendation incorporating user relationships,we make a study on the personalized conference paper recommendation in academic social media by incorporating the academic relationships among users and propose a unified recommendation model which takes a paper's content information and the academic relationships among users and the authors of a paper into consideration.The model exploits three different kinds of academic relationships between a user and the authors of a paper for recommendation:citation relationships,coauthor relationships,and research interest correlations.In the model,given a user and a paper,we extract several features from three different types of information resources:the citation network,the coauthor network,and contents,respectively.In addition,we propose to use a user's bookmarks in a conference to indicate the user's pairwise preferences towards all the papers in that conference.Furthermore,we employ a pairwise learning to rank model which exploits the pairwise user preference to learn a function that predicts a user's preference towards a paper based on the extracted features.The experimental results on real datasets demonstrate that our proposed recommendation model is effective in recommending conference papers.(3)For the problem of personalized social relation recommendation exploiting the heterogeneous network,we make a study on the personalized followee recommendation in event-based social networks by exploiting the heterogeneous nature of an event-based social network and propose a unified recommendation model which considers both the online social network and the offline event participation network of an event-based social network.In the model,we extract explicit features and latent features from the online social network and the offline event participation network of an event-based social network,respectively.Also,we propose a similarity regularization method to put a constraint on a user's latent features,which forces a user's latent features close to his/her followees'.Besides,based on users' pairwise preferences extracted from users' online social relations and event participation records,we propose a novel Bayesian optimization framework which takes AUC as the optimization object.The experimental results on real datasets demonstrate that our proposed recommendation model is effective in recommending followees in event-based social networks.