Distributed methods for large scale regression

Automatically collected data in environmental modelling, energy management and medicine may involve very large data volumes while also requiring richly parameterized models for adequate analysis and prediction. If n is data set size and p the number of model coefficients, this project aims to find O(np) computational methods for estimating penalized regression models, which are susceptible to parallelization in cluster computing environments. The major challenge is to do this in a way that adequately estimates hyper-parameters alongside regression coefficients, and the project will investigate the feasibility of doing this using stochastic log determinant or log trace estimators in the context of marginal likelihood or similar criteria.