Table 4. Median Absolute Error (MAE) of the data reconstruction from different manifold learning embeddings.

MAE is assessed using Leave-One-Out procedure. The reconstruction is done by the independent k-nearest neighbors regression of the coordinates in the original space of relative taxon abundances from the non-linear embedding. Notation as in Table 2.

Dataset	Method	Tax O	Tax F	Tax G
AGP	AutoEncoder	0.06	0.19	0.22
	t-SNE	0.05	0.20	0.22
	UMAP	0.06	0.22	0.24
	Isomap	0.06	0.22	0.25
	LLE	0.06	0.21	0.24
	Spectral	0.06	0.21	0.24
HMP	AutoEncoder	0.09	0.24	0.22
	t-SNE	0.09	0.24	0.22
	UMAP	0.11	0.27	0.25
	Isomap	0.11	0.29	0.27
	LLE	0.12	0.29	0.26
	Spectral	0.13	0.34	0.29