UCLA

The problem of information network analysis has gained increasing attention in recent years, because most objects and data in the real world are interconnected, forming complex networks. One challenging aspect of network analysis is that they are inherently resistant to parametric modeling, which allows us to truly express the vertices and edges in the network as vectors or functions of time. This is because, unlike multi-dimensional data, the edges in the network reflect interactions among vertices, and it is difficult to independently model them without taking into account their correlations and interactions with neighboring vertices or edges. This thesis presents a combination of the methods and applications in static network analysis and evolutionary network analysis, which is trying to analyze and model the structure and evolution in networks.

On the analysis of static network side, we develop methods to learn the network representations with adversarially regularized autoencoders, which learns smoothly regularized vertex representations that well capture the network structure through jointly considering both locality-preserving and global reconstruction constraints. And applications in community detection are designed to detect communities which have the maximum competition intensity score in an advertiser-keyword bipartite network.

On the analysis of evolutionary network side, we show that it is indeed possible to represent the network structure purely as a function of time with the use of temporal matrix factorization, in which the entries are parameterized by time. This opens the possibility of using the approach for a wide variety of evolutionary network analysis problems, such as temporal link prediction. It can also be utilized to model co-evolution across multiple networks by decomposing the adjacency matrix of each co-evolving network into a product of network-independent factors and a set of network-specific time-dependent factors. Applications in temporal link prediction and online anomaly detection are proposed, in which the low-dimensional representations of networks can be learned and updated to capture evolutionary network structures.