B1303
Title: Two generalizable strategies for scalable inference from network data
Authors: Srijan Sengupta - North Carolina State University (United States) [presenting]
Abstract: Massive network data are becoming increasingly common in scientific applications. Existing community detection methods are computationally infeasible for such massive networks. Two generalizable strategies are proposed for scalable inference from network data: SONNET and predictive subsampling. SONNET is a divide-and-conquer algorithm where the original network is split into multiple subnetworks with a common overlap. Statistical inference is carried out for each subnetwork, and the results from individual subnetworks are aggregated by leveraging the overlap. The core idea of predictive subsampling is to avoid large-scale matrix computations by breaking up the task into a smaller matrix computation plus a large number of vector computations that can be carried out in parallel. Under the proposed method, the inferential task of interest is carried out on a small subgraph to estimate the relevant model parameters. The remaining nodes are added one by one using only vector computations. These two strategies are applied to various inference tasks, such as community detection, parameter estimation, model selection, and hypothesis testing.