Sparse principal balances for high-dimensional compositional data

Abstract

Extracting the most essential information out of compositional data can be done by a method called principal balances [5]. This method is, however, compu- tationally only feasible for low-dimensional data. For high-dimensional composi- tional data we introduce the concept of sparse principal balances, a method that relies on sparse principal component analysis to construct principal directions with many zero loadings. Sparse principal balances are fast to compute even for very high-dimensional data, and their interpretation is easier than principal balances.