Scalable learning of users' preferences using networked data

Users' personal information such as their political views is important for many applications such as targeted advertisements or real-time monitoring of political opinions. Huge amounts of data generated by social media users present opportunities and challenges to study these preferences in a large scale. In this paper, we aim to infer social media users' political views when only network information is available. In particular, given personal preferences about some of the social media users, how can we infer the preferences of unobserved individuals in the same network? There are many existing solutions that address the problem of classification with networked data problem. However, networks in social media normally involve millions and even hundreds of millions of nodes, which make the scalability an important problem in inferring personal preferences in social media. To address the scalability issue, we use social influence theory to construct new features based on a combination of local and global structures of the network. Then we use these features to train classifiers and predict users' preferences. Due to the size of real-world social networks, using the entire network information is inefficient and not practical in many cases. By extracting local social dimensions, we present an efficient and scalable solution. Further, by capturing the network's global pattern, the proposed solution, balances the performance requirement between accuracy and efficiency.