A Multicriteria Approach to Find Predictive and Sparse Models with Stable Feature Selection for High-Dimensional Data

Table 5

Pareto optimal configurations in Figure 9 whose resulting models averagely use less than 500 features (above the horizontal line) and accuracy optimal configurations (below the horizontal line). The ID column references Table 6.

ID	Classifier	Parameters	Filter	n.feats	Error	Error	Size	Size
					(train)	(test)	(train)	(test)
1	Random Forest	num.trees = 20577,	Variance	397	0.041	0.048	397.0	397.0
		min.node.size = 2
2	SVM	,	Variance	355	0.040	0.044	355.0	355.0
3	Lasso Log. Reg.		Variance	246	0.041	0.059	246.0	245.9
4	Lasso Log. Reg.		Variance	232	0.037	0.044	232.0	232.0
5	Random Forest	num.trees = 4742,	Variance	212	0.048	0.044	212.0	212.0
		min.node.size = 2
6	Lasso Log. Reg.		Variance	174	0.044	0.048	174.0	174.0
7	GLM Boosting		Variance	728	0.041	0.041	13.3	15.1
8	GLM Boosting		Variance	363	0.048	0.074	9.8	7.2
9	GLM Boosting		AUC	232	0.048	0.077	9.7	11.0
10	Random Forest	num.trees = 1062,	MRMR	1756	0	0.055	1466.0	1611.9
		min.node.size = 3
11	Random Forest	num.trees = 4671,	AUC	760	0	0.051	759.7	760.0
		min.node.size = 11