PhD Qualifier Written Critiques

Due to the availability of diverse information, the online social networks are of much heterogeneity, which is formally defined as Heterogeneous Information Network (HIN). Meanwhile, users usually participate in multiple networks simultaneously, but not all of them are well-labeled. By transferring the information from the well-labeled source network to target network that is lack of label information, we can solve the information insufficiency problem in the target network. However, this problem is non-trivial since:(1) the distributions of different networks are different;(2) the correlation of different networks are hard to model;(3) information in HIN is hard to retrieve. In this paper, we review the methods related to transfer learning in HIN from three perspectives. The first is the transfer learning in bipartite HIN, which is the relational learning with matrix factorization methods. The second is to investigate the transferable features in general HINs with multiple types of nodes, which is the link-based features based on social theories. The last one is to leverage the meta-path based transferable features, which extract the high-order semantics of HIN both within networks and across networks. We also present some possible future directions related to transfer learning in HIN.