Ehancing Geo-Social Systems: Profiling, Ranking and Recommendation

The global sharing of fine-grained geo-spatial footprints – via smart mobile devices and social media services (e.g., Facebook, Twitter, Flickr) – is leading to the creation of a new class of geo-social systems. These systems promise new insights into the dynamics of human behavior and new intelligent location-aware applications, enabled by both geographic and social characteristics of their users. Yet, we are just beginning to understand these nascent systems. Indeed, there is a significant research gap in understanding, modeling, and leveraging geo-tagged user activities and in identifying the key factors that influence the success of new geo-social systems. Hence, this dissertation research seeks to enhance existing and future geo-social systems through a systematic study of the intrinsic mutual reinforcement relationship between geography (geo) and user behaviors (social). In particular, we focus on three important scenarios: geo-social profiling, ranking, and recommendation. In summary, this dissertation makes three unique contributions: The first contribution of this dissertation research lies in frameworks for profiling both users and locations in geo-social systems. We propose a framework to identify conceptual communities and a smoothing-based approach that collectively balances the information from physical neighbors and conceptual community for estimating the hashtag distribution at a particular location. We further propose a location-sensitive folksonomy construction framework and build high-quality tag profiles for users by identifying candidate tags from these location-sensitive folksonomies, and then employ a learningto- rank approach for ordering these tags. The second contribution of this dissertation research is a framework for location-sensitive topical expert identification and ranking in social media. Three of the key features ii of the proposed approach are: (i) a learning-based framework for integrating multiple user-based, content-based, list-based, and crowd-based factors impacting local expertise that leverages the fine-grained GPS coordinates of millions of social media users; (ii) a location-sensitive random walk that propagates crowd knowledge of a candidate’s expertise; and (iii) a comprehensive controlled study over AMT-labeled local experts on eight topics and in four cities. In the third contribution, we create a new geo-social framework for effective personalized image recommendation as part of the effort toward enhancing geo-social systems. We propose Neural Personalized Ranking (NPR) – a personalized pairwise ranking model over implicit feedback datasets – that is inspired by Bayesian Personalized Ranking (BPR) and recent advances in neural networks. We further build an enhanced model which significantly boosts performance by augmenting the basic NPR model with multiple contextual preference clues derived from geographic features, user tags and visual factors