People are interconnected through online social networks ubiquitous nowadays. The analysis of these networks also attracts many research interests with a broad range of applications. Various studies have been presented to study the network structure as well as users’ social characteristics. Despite of their success, most previous research works focus on analyzing individual networks. However, data in individual networks can be quite sparse and each individual social network may reflect only partial aspects of users’ social behaviors. Building models on such networks may overfit the rare observations and fail to capture the whole picture of users’ social interests. In reality, nowadays people join multiple networks for different purposes. For example, users may use Facebook to co...[ Read more ]

People are interconnected through online social networks ubiquitous nowadays. The analysis of these networks also attracts many research interests with a broad range of applications. Various studies have been presented to study the network structure as well as users’ social characteristics. Despite of their success, most previous research works focus on analyzing individual networks. However, data in individual networks can be quite sparse and each individual social network may reflect only partial aspects of users’ social behaviors. Building models on such networks may overfit the rare observations and fail to capture the whole picture of users’ social interests. In reality, nowadays people join multiple networks for different purposes. For example, users may use Facebook to connect with their friends, talk with their families on Skype and follow celebrities on Twitter, etc. Thus, different networks are correlated with each other and nested together as composite social networks by the shared users. If we consider these users as the bridge, fragmented knowledge in individual networks can be utilized collectively to build more accurate models and obtain comprehensive understandings of users’ social behaviors.

In this research, our main idea is to extract common knowledge from different networks to solve the data sparsity problem but takes care of the network differences. We propose a general framework, known as ComSoc, based on hierarchical Bayesian models, by encoding common knowledge and network differences as latent factors. Based on this framework, we analyze composite social networks from four major aspects: 1). how to model the composite network structures; 2). how to model the dynamics and network co-evolution; 3). how to adaptively predict users’ social behaviors across social medias; and 4). how to measure users’ distances specifically in different networks. We will use large-scale social networking datasets, to carry out this research, in order to demonstrate how our ComSoc framework can be instantiated for solving these four problems. Finally, to handle big data, we propose a novel parallel framework that makes the model inference efficient.