Key Technologies For Spatio-temporal Recommendation On Social Media

With the development of information technology and the Internet, the amountof information available increases exponentially, leading to the “information explo-sion” and “information overload”. Moreover, with the rise of social media, suchas social networking service(SNS), location-based social networking service (LBSN)and mobile social media (e.g., weixin), the ecosystem on which search engines relyfor existence has been fundamentally shaken. For one thing, most of the informationon social media is no longer open and free; for another, the information itself is grad-ually being replaced by user experience. Besides, the online data on social mediaand people’s ofine activities have gradually become integrated. The conventionalsearch technology is unable to efectively integrate the personal online private datawith people’s ofine activities, while this trend provides an unprecedented marketopportunities and application scenarios for the recommender system as social mediamarketing has been widely adopted by more and more businesses. For one thing,the recommender system can help users make choices by providing users interest-ing items and services; for another, the recommender system can help the businesspush their items to the right users. The win-win situation can be achieved by therecommender system. With the advancement of machine learning and data mining technology, thedevelopment of the recommender system has been successful in both research andapplication. However, due to lack of deep understanding and modeling of people’sbehaviors in the spatio-temporal context, there are still a number of challengingresearch problems on temporal or/and spatial recommendation for social media,such as “temporal context-aware recommendation”,“recommendation for out-of-town users” and “location-based and time-aware recommendation”. To addressthese challenges, this dissertation exploits time information, location informationand spatio-temporal information for user modeling on social media. The main con-tributions of this dissertation are summarized as follows:Temporal Context-Aware Recommendation: Social media provides valuable re-sources to analyze user behaviors and capture user preferences. We analyzeduser behaviors in social media systems and designed a latent class statisticalmixture model, named temporal context-aware mixture model (TCAM), toaccount for the intentions and preferences behind user behaviors. Based onthe observation that the behaviors of a user in social media systems are gen-erally infuenced by intrinsic interest as well as the temporal context (e.g.,the public’s attention at that time), TCAM simultaneously models the topicsrelated to users’ intrinsic interests and the topics related to temporal contextand then combines the infuences from the two factors to model user behaviorsin a unifed way. The experimental studies demonstrated the superiority ofthe proposed TCAM models on four popular social media datasets in the tasksof temporal recommendation, user behavior analysis and topic detection.Recommendation for Out-of-Town Users: Newly emerging location-based andevent-based social network services provide us with a new platform to un-derstand users’ preferences based on their activity history. A user can onlyvisit a limited number of venues/events and most of them are within a lim-ited distance range, so the user-item matrix is very sparse, which creates abig challenge to the traditional collaborative fltering-based recommender sys-tems. The problem becomes even more challenging when people travel out of town, especially in a new city where they have no activity information.To address this problem, we proposed a location-content-aware recommendersystem, LCARS, which provides a user with spatial item recommendationswithin the querying city based on the individual interest and the local pref-erence mined from the user’s activity history. LCARS can facilitate people’stravel not only near their living areas but also to a city that is out of towneven if they do not have any activity history there. The experimental result-s show that our approach signifcantly outperforms existing recommendationmethods in the efectiveness of both home-town and out-of-town recommenda-tion. Meanwhile, the proposed scalable query processing technique based onTA algorithm, improves the efciency of our approach signifcantly, enablinga fast online recommendation scenario.Location-based and Time-Aware Recommendation: The unique properties oflocation based social networks (LBSN) pose a great challenge for the usercheck-in behavior modeling task. In LBSN, the decision process of a userchoosing a POI is complex and can be infuenced by multiple factors, such asgeographical infuences, user interests, temporal efects and local attractions.While a growing line of research has focused on modeling user check-in be-haviors, it lacks of integrated analysis of the joint efect of the various factors.To this end, we proposed a joint probabilistic generative model to accountfor user check-in behaviors on location-based social networks (LBSN). Theproposed model strategically takes into consideration various factors whichinfuence the user check-in decision process. Then, we deployed our modelto location-based and time-aware POI recommendation, i.e., to recommendPOIs for a given user at a specifc time and location. Finally, experimentalresults on two real-world datasets validated the superiority of the proposedmodel.