UCLA

Discovering the latent social characteristics of nodes in an information network has been an important task. The latent merits and characteristics can be represented as a vector, where each dimension may be associated with a special meaning (e.g. topic). These vector representations are essential in understanding the users and the network, and they are applicable to many network-based tasks as well, such as clustering, classification, and link prediction and so on.

In this thesis, I will start from the political ideology estimation problem as a valid example of latent social characteristics. I will cover learning users' latent characteristics (embedding) through (1) bipartite networks between user and other entities; (2) networks with different types of links between users and (3) evolving social networks. Extensive evaluation and case studies are conducted to demonstrate the effectiveness of the estimated representations. Finally, I will investigate the scenario where multiple sets of social characteristics are present in the network, and generalize the methodology into a boarder and more general framework.