Predictive modeling for navigating social media

Social media changes the way people use the Web. It has transformed ordinary Web users from information consumers to content contributors. One popular form of content contribution is social tagging, in which users assign tags to Web resources. By the collective efforts of the social tagging community, a new information space has been created for information navigation. Navigation allows serendipitous discovery of information by examining the information objects linked to one another in the social tagging space. In this dissertation, we study prediction tasks that facilitate navigation in social tagging systems. For social tagging systems to meet complex navigation needs of users, two issues are fundamental, namely link sparseness and object selection. Link sparseness is observed for many resources that are untagged or inadequately tagged, hindering navigation to the resources. Object selection is concerned when there are a large number of information objects that are linked to the current object, requiring to select the more interesting or relevant ones for guiding navigation effectively. This dissertation focuses on three dimensions, namely the semantic, social and temporal dimensions, to address link sparseness and object selection. To address link sparseness, we study the task of tag prediction. This task aims to enrich tags for the untagged or inadequately tagged resources, such that the predicted tags can serve as navigable links to these resources. For this task, we take a topic modeling approach to exploit the latent semantic relationships between resource content and tags. To address object selection, we study the task of personalized tag recommendation and trend discovery using social annotations. Personalized tag recommendation leverages the collective wisdom from the social tagging community to recommend tags that are semantically relevant to the target resource, while being tailored to the tagging preferences of individual users. For this task, we propose a probabilistic framework which leverages the implicit social links between like-minded users, i.e. who show similar tagging preferences, to recommend suitable tags. Social tags capture the interest of the users in the annotated resources at different times. These social annotations allow us to construct temporal profiles for the annotated resources. By analyzing these temporal profiles, we unveil the non-trivial temporal trends of the annotated resources, which provide novel metrics for selecting relevant and interesting resources for guiding navigation. For trend discovery using social annotations, we propose a trend discovery process which enables us to analyze trends for a multitude of semantics encapsulated in the temporal profiles of the annotated resources.