Fig. 14
From: Modeling microbiome-trait associations with taxonomy-adaptive neural networks
Flattening the internal taxonomic units as additional features does not effectively exploit the taxonomy. The feature value of each internal taxonomic unit is computed as the average of the values across all taxa within that specific taxonomic unit. Each tree-agnostic model takes as input the concatenated features from both the taxa and the internal taxonomic units. Treating the internal taxonomic units as additional features results in marginal or even worse predictive performance, as measured by both AUPRC and AUROC