compositional-data

This paper critiques standard attention metric methods in transformers, revealing that choices like keeping or dropping sink tokens can reverse conclusions, and proposes using compositional data analysis to separate sink and content attention components for more accurate interpretation.

This paper introduces PolyILR, a canonical orthonormal decomposition of the Aitchison tangent space that aligns with any tree topology, providing stable and interpretable features for compositional data such as microbiome and single-cell profiles.