Three approaches to supervised learning for compositional data with pairwise logratios

The common approach to compositional data analysis is transform the by means of logratios. Logratios between pairs parts (pairwise logratios) are easiest interpret in many research problems. When number large, some form logratio selection a must, for instance an unsupervised learning method based on stepwise pairwise logratios that explain largest percentage variance dataset. In this article we present three alternative supervised methods select best dependent variable generalized linear model, each geared specific problem. first features unrestricted search, where any can be selected. This has complex interpretation if overlap, but it leads most accurate predictions. second restricts occur only once, which makes corresponding intuitively interpretable. third uses additive logratios, so $K-1$ selected involve exactly $K$ parts. fact searches subcomposition with highest explanatory power. Once identified, researcher's favourite representation may used subsequent analyses, not Our methodology allows or non-compositional covariates forced into models theoretical knowledge, and various stopping criteria available information measures statistical significance Bonferroni correction. We illustration approaches dataset from study predicting Crohn's disease. excels terms predictive power, other two interpretability.