Table 1.

Cluster tightness and separation (Davies-Bouldin, Calinski-Harabaz, Silhouette), as well as subclass recovery performance (Adjusted Mutual Information after hierarchical clustering), in explanation space (e) and feature space (f), in original dimensionality and post PCA, Isomap, and T-SNE. Higher values are better (except for Davies-Bouldin, where lower is better).

Data	Dim. Red.	Space	Davies-Bouldin	Calinski-Harabaz	Silhouette	Adjusted MI
Sythetic	None	(f)	7.00	6.52	0.02	0.08
		(e)	1.11	192.62	0.46	0.59
	PCA	(f)	3.30	38.08	0.01	0.14
		(e)	0.78	258.24	0.66	0.59
	Isomap	(f)	12.90	3.93	-0.02	0.02
		(e)	0.72	187.73	0.71	0.52
	T-SNE	(f)	29.45	0.61	-0.02	0.00
		(e)	0.29	2937.57	0.79	0.97
Fashion	None	(f)	11.10	2.86	-0.03	0.00
		(e)	1.32	205.45	0.30	0.61
	PCA	(f)	5.72	2.18	-0.12	0.00
		(e)	0.74	581.87	0.51	0.56
	Isomap	(f)	4.87	75.60	-0.07	0.00
		(e)	0.92	398.96	0.41	0.56
	T-SNE	(f)	2.30	65.03	0.12	0.13
		(e)	0.69	768.13	0.52	0.71
UKBB	None	(f)	4.24	26.38	0.04	0.04
		(e)	2.33	84.91	0.12	0.15
	PCA	(f)	2.53	57.65	0.07	0.05
		(e)	1.54	152.25	0.21	0.14
	Isomap	(f)	2.60	60.75	0.08	0.04
		(e)	1.74	122.90	0.21	0.14
	T-SNE	(f)	2.64	59.99	0.08	0.05
		(e)	1.44	180.59	0.24	0.12
TCGA	None	(f)	8.58	8.54	0.01	0.10
		(e)	2.64	57.37	0.13	0.68
	PCA	(f)	4.44	17.81	-0.01	0.05
		(e)	0.59	929.48	0.59	0.67
	Isomap	(f)	8.27	70.70	0.02	0.16
		(e)	0.51	834.70	0.62	0.70
	T-SNE	(f)	5.03	106.36	0.11	0.28
		(e)	0.51	1561.08	0.60	0.73