Table 3.

Accuracy of the PD/NC classification, compared among baseline classifiers and different feature-sample selection or reduction techniques. First column shows the results for the proposed joint feature-sample selection method. The second, third and fourth columns include the results with separate feature and sample selection, sparse feature selection, and no feature or sample selections, respectively. The next five columns show the results for some state-of-the-art feature reduction techniques, and finally the last column shows the results for the well-known RANSAC algorithm for outlier sample removal.

Classifier	Selection/Reduction method
	JFSS	FSS	SFS	no FSS	mRMR	PCA	RPCA	AE-RBM	NNMF	RANSAC
Robust LDA	81.9	78.0*	67.7†	61.5†	70.5†	65.0†	N/A	76.8*	64.5†	74.7†
MC	78.9*	73.5	66.0†	56.2†	69.2†	62.4†	N/A	73.1†	64.1†	72.3†
LDA	65.9†	62.1†	61.5†	56.0†	60.9†	56.0†	60.5†	65.1†	58.1†	66.0†
SVM	69.1†	61.9†	61.1†	55.5†	58.8†	58.5†	61.0†	66.6†	59.1†	71.2†
Sparse SVM	70.1†	62.8†	6.15†	59.5†	60.0†	59.3†	61.8†	68.7†	63.1†	73.1†
SR	N/A	61.6†	59.6†	N/A	60.5†	59.9†	60.6†	63.7†	61.5†	64.2†
JFSS-C	68.7†	N/A	N/A	N/A	67.5†	68.8†	71.9†	72.8†	67.0†	69.9†

Note that

^*stands for the case with p<0.05 and

^†for p<0.01 in a cross-validated 5×2 t-test against the proposed method (RLDA + JFSS). Bold indicates best achieved results.