Structured Regression on Multiscale Networks

Abstract

Structure-based regression algorithms generally suffer substantive speed losses and have exacting memory requirements compared to their structureless counterparts. Gaussian conditional random field (GCRF) models are one of the most time- and memory-efficient approaches to structured regression. The authors' previous speedups for the GCRF method allow for exact solutions on networks of up to 100,000 nodes and 10 million links. Using multiscale networks, the exact solution for networks of millions of nodes and trillions of links can be solved in a similar amount of time. They walk through the intuitiveness of using multiple scales of networks on a real-life health informatics application. The time and memory demands from using this approach are logarithmic compared to naive implementations.