A1584
Title: Representation learning of dynamic networks
Authors: Haixu Wang - University of Calgary (Canada) [presenting]
Abstract: A representation learning model for dynamic networks, which describes the continuously changing relationship between individuals in a population, is established. The problem is encapsulated in a dimension reduction problem of functional data analysis. Given that the dynamic networks are a form of matrix-valued functions, the goal is to find a mapping that compresses such functional data into vector-valued functions. The learning space, which is a much lower dimensional function space, is endowed with norm and inner product. Moreover, the learning method is further allowed to be asymmetric in the learning space so that the in and out roles of individual nodes can be studied separately. In simulation studies, the finite sample performance of the method in link prediction is compared to the existing methods. Sociological networks are common examples in real data applications. The representation learning method is applied to the social networks of six ant colonies where network connections are observed over a time span of 41 days. The interactions with colonies are well captured in the low-dimensional
