Robust Structure Preserving Nonnegative Matrix Factorization for Dimensionality Reduction

Table 3

Clustering performance on local noise contaminated datasets. The best performances are highlighted in bold font.

(a)

Dataset	Accuracy (%)
	means	PCA	NMF	GNMF	SPNMF	RSPNMF
FERET	23.10 ± 1.14	26.12 ± 1.27	24.20 ± 1.03	25.05 ± 0.74	27.45 ± 5.22	33.06 ± 1.25
ORL	35.71 ± 2.74	46.99 ± 2.69	44.56 ± 1.59	37.78 ± 1.37	44.78 ± 1.7	54.01 ± 3.3
Yale	32.12 ± 2.38	37.43 ± 2.27	38.18 ± 2.18	32.45 ± 1.16	36.36 ± 1.24	44.42 ± 1.08

(b)

Dataset	Normalized mutual information (%)
	means	PCA	NMF	GNMF	SPNMF	RSPNMF
FERET	49.2 ± 2.56	56.85 ± 1.54	54.33 ± 1.12	50.06 ± 2.51	58.26 ± 0.89	67.44 ± 1.73
ORL	53.78 ± 2.56	65.79 ± 1.54	64.57 ± 1.12	58.65 ± 2.51	64.26 ± 0.89	71.50 ± 1.73
Yale	35.73 ± 2.49	42.72 ± 2.30	43.77 ± 1.58	38.84 ± 0.98	42.72 ± 1.12	47.56 ± 1.19