Effects of Pooling Samples on the Performance of Classification Algorithms: A Comparative Study

Table 1

Comparison of classification performance with different numbers of top-ranked features.

	Misclassification rate
Classifiers	Top 10	Top 100	Top 1000
Individual
SVM with linear kernel	0.2191**	0.3479**	0.3757
SVM with radial kernel	0.2178**	0.3211**	0.3712
RF	0.2436**	0.2926**	0.2975
𝑘-NN	0.2515**	0.3270**	0.4354
PLR	0.2160**	0.3329**	0.3761
PAM	0.2096**	0.3185**	0.2310
Pool size = 5
SVM with linear kernel	0.3229**	0.3817**	0.4131
SVM with radial kernel	0.3167**	0.3771**	0.4571
RF	0.3272**	0.3568**	0.3841
𝑘-NN	0.3133**	0.3779**	0.4720
PLR	0.3113**	0.3772**	0.3983
PAM	0.3005**	0.3799*	0.3681

Classification performance is presented for different number of top-ranked features. The dataset contains a total of 1000 features and 90 samples per class with 10 markers. Top 1000 features denote no feature selection. The table shows results using individual samples and illustrates results derived by means of a pooled dataset when pool size is 5, respectively. Significance levels ^∗𝑃<0.05 and ^∗∗𝑃≪0.05 indicate comparisons where no feature selection is performed by using the Wilcoxon rank sum test.