Three approaches to supervised learning for compositional data with pairwise logratios.

Area of Science:

• Statistics
• Data Science
• Bioinformatics

Background:

• Compositional data analysis (CoDa) often involves high-dimensional data.
• Pairwise logratios are key for interpreting CoDa, but selection is challenging with many parts.
• Generalized linear models (GLMs) are frequently used for analyzing such data.

Purpose of the Study:

• To develop and present three novel stepwise supervised learning methods for selecting optimal pairwise logratios.
• To enhance the interpretability and predictive accuracy of GLMs in CoDa.
• To provide flexible model-building strategies accommodating prior knowledge and various stopping criteria.

Main Methods:

• Three stepwise supervised learning approaches for logratio selection: unrestricted search, restricted search (unique parts), and additive logratios.
• Integration of logratios or covariates into GLMs, with options for forced inclusion.
• Application of information criteria or Bonferroni-corrected statistical significance for model selection stopping rules.

Main Results:

• The unrestricted search method yields the highest prediction accuracy, though interpretation can be complex.
• The restricted search method offers more intuitive interpretability by ensuring unique part usage in logratios.
• The additive logratio method facilitates the analysis of subcompositions.

Conclusions:

• The proposed methods offer effective strategies for selecting informative logratios in CoDa for GLMs.
• The choice of method depends on the balance between predictive performance and interpretability.
• The application demonstrates the utility of these methods in a real-world biomedical study (Crohn's disease prediction).