Penalized polytomous ordinal logistic regression using cumulative logits. Application to network inference of zero-inflated variables

TL;DR

This paper introduces a penalized ordinal logistic regression method with stability selection and knockoffs for variable selection, applied to network inference of zero-inflated variables in real agronomic data.

Contribution

It develops a novel penalized regression approach using cumulative logits, stability selection, and knockoffs for ordinal responses, with practical application to zero-inflated network inference.

Findings

Effective variable selection in ordinal regression models.

Successful application to real agronomic zero-inflated data.

Enhanced interpretability of covariate importance.

Abstract

We consider the problem of variable selection when the response is ordinal, that is an ordered categorical variable. In particular, we are interested in selecting quantitative explanatory variables linked with the ordinal response variable and we want to determine which predictors are relevant. In this framework, we choose to use the polytomous ordinal logistic regression model using cumulative logits which generalizes the logistic regression. We then introduce the Lasso estimation of the regression coefficients using the Frank-Wolfe algorithm. To deal with the choice of the penalty parameter, we use the stability selection method and we develop a new method based on the knockoffs idea. This knockoffs method is general and suitable to any regression and besides, gives an order of importance of the covariates. Finally, we provide some experimental results to corroborate our method. We then present an application of this regression method for network inference of zero-inflated variables and use it in practice on real abundance data in an agronomic context.