Study Of Recommendation Models,User Loyalty And Group Activeness In Event-based Social Networks

Event-based social networks(EBSNs)are a new type of social platforms that integrate online social communities with offline social activities.Meetup and Douban are two representative examples which have become popular recently.These social networks provide services for various groups of users who have common interests in both online activities and offline events.EBSN users create or join different user groups for social or professional networking,such as joining some online social activities or attending offline events organized at different times and locations.Since a large number of users have joined these event-based social networks and various online and offline social events and activities are organized through the networks,tremendous social networking data is generated in EBSNs.This data provides a new golden mine for study of event-based social networks and also poses new research problems in social network analysis,for instance,recommendation problems,interest group stability problems and user loyalty problems.In this thesis,we study the following three research problems in event-based social networks.The first research problem is about event attendees recommendation,i.e.,recommending an upcoming event to a list of attendees based on the previous events these attendees attended.To tackle this problem,we propose a new location-based topic model which is built of the information of social event topics,and the locations and time of the social events.In this new model,we convert the information of events into topic probability vectors using the Latent Dirichlet Allocation(LDA)topic model and Gibbs sampling method.The set of previous events whose themes are similar to the upcoming event is discovered based on the similarities between the previous events and the upcoming event.The users who attended the previous events are retrieved from the EBSN data and scored according to the locations,time,similarity of the chosen events to the upcoming event.Finally,a list of users with high scores is selected as candidates to whom the upcoming event is recommended.The experiment analysis shows that the new recommendation model has better performance over the content-based and location-based recommendation models.The second research problem is a new recommendation problem we have defined in this thesis,i.e.,recommending an upcoming event to the top N active friends of a key user who creates that event or will attend that event.This problem is different from the previous problem in that it targets at not only the users who are likely to attend an upcoming event but also the active friends of these users.The active friends may not attend the previous events similar to the upcoming event.Therefore,the upcoming event is recommended to a broader group of users.To solve this problem,the interaction of social event topics and personal behaviors of users are investigated to understand the decisions of users whether to engage in events or not.The contents of events and the behaviour of users are exploited to construct an association matrix of events and users.We propose a new content-based event recommendation model that finds a list of active friends to the key user from the association matrix and uses the score model to score active friends based on interest and behavior similarities as well as time difference.The new recommendation model finally recommends the upcoming event to the top N friends with high scores.This model is a double user level recommendation model while the event attendees recommendation one is a single user level recommendation which is used in most existing recommendation models.In the third research problem,we investigate the activeness and loyalty of EBSNs users and give a novel understanding of the growth and the inactiveness of social groups within a particular time frame.The entities and structures of the networks are analyzed to generate three sets of features,i.e.,group-based,event-based and user-based features.The concepts of user loyalty and group activeness are defined within a series of consecutive time-windows.Each window presents the activities of groups and users in their groups.To measure the activeness,we propose a new method that is based on the numbers of events in the given time-windows.In this method,the fraction of events of a group between two consecutive windows is first calculated,and then an association matrix is created to assign one suitable activeness label to the corresponding group within several time windows.Similarly,the user loyalty is measured by the attended events within those windows and is then considered as an important feature to the group activeness.We select three prevalent classifiers to validate the activeness labels of groups.As a result,the small group of correlated features is found,and these features predict higher accuracy compared to all features.These three problems are very important in understanding and maintaining event-based social networks.They are also relevant to other social networks.Although the first recommendation problem is well known in other social networks,we study it in the event-based social network perspective and target it to the users who are likely to attend an offline event.Therefore,location and time of the event have to be considered for travel convenience.The last two problems are newly defined in this thesis and the solutions,which are evaluated using the real data,are effective and innovative.The research results in this thesis contribute to new methods and tools for analysis of event-based social networks.