Dominating Hyperplane Regularization for Variable Selection in Multivariate Count Regression

Abstract

Identifying relevant factors that influence the multinomial counts in compositional data is difficult in high dimensional settings due to the complex associations and overdispersion. Multivariate count models such as the Dirichlet-multinomial (DM), negative multinomial, and generalized DM accommodate overdispersion but are difficult to optimize due to their non-concave likelihood functions. Further, for the class of regression models that associate covariates to the multivariate count outcomes, variable selection becomes necessary as the number of potentially relevant factors becomes large. The sparse group lasso (SGL) is a natural choice for regularizing these models. Motivated by understanding the associations between water quality and benthic macroinvertebrate compositions in Canada's Athabasca oil sands region, we develop dominating hyperplane regularization (DHR), a novel method for optimizing regularized regression models with the SGL penalty. Under the majorization-minimization framework, we show that applying DHR to a SGL penalty gives rise to a surrogate function that can be expressed as a weighted ridge penalty. Consequently, we prove that for multivariate count regression models with the SGL penalty, the optimization leads to an iteratively reweighted Poisson ridge regression. We demonstrate stable optimization and high performance of our algorithm through simulation and real world application to benthic macroinvertebrate compositions.