On recommendations in heterogeneous social media networks

In this dissertation, we study the problem of social media recommendations with a heavy emphasis on exploiting social, content and contextual information. The problem of recommendation analysis and collaborative filtering has been widely studied in the literature because of its numerous applications to a wide variety of scenarios. Many social media sites such as Flickr or YouTube contain multimedia objects, which occur in the context of an extensive amount of content information such as tags, image content, in addition to the user preferences for the different objects. Thus, there is a plethora of heterogeneous, content, linkage and preference information in a social media network, which can be used in order to make effective recommendations in such networks. In this dissertation, we will study the problem of making recommendations in such complex multimedia networks with the use of such information. While our approach is developed and evaluated for the case of the Flickr image network, the broad principles are applicable to any kind of multimedia network such as a music or video site. To ensure the efficiency and scalability, we further extend our approach to incorporate the latent factor model so that our approach can be very useful for making personalized content recommendations in large and heterogeneous social media networks. This dissertation also studies a variety of recommendation scenarios, including context-specific recommendations which are made based on some kinds of content the user specifies, cold start recommendations that utilizes the social relations to make recommendations when a new user gets on board, and preference drifting to realize the long-term change of users' tastes. We present experimental results illustrating the effectiveness and efficiency of our approaches