Table 2.

Descriptions of the 13 feature selection methods in this study.

Type	Method	Description	Equation
FITI	MIM	Evaluates features by correlation between features and classes measured by mutual information	$MIM(f_i)=I(f_i;C)$
	MIFS/MRMR	Evaluates features by correlation between features and classes and redundancy among features	$MIFS(f_i)=I(f_i;C)-\beta {\displaystyle \sum _{s_j\in S}I(f_i;f_S)}$ $MRMR(f_i)=I(f_i;C)-\frac{1}{S}{\displaystyle \sum _{s_j\in S}I(f_i;f_S)}$
	JMI/CMIM	Evaluates features by correlation between features and classes and redundancy among features measured by conditional mutual information	$JMI(f_i)=I(f_i;C)-\frac{1}{\left|S\right|}{\displaystyle \sum _{s_j\in S}[I(f_i;C)-I(f_i;C|f_S)}]$ $CMIM(f_i)=\underset{f_s\in S}{\min }I(f_i;C|f_s)$
SIF	Fisher/LS	Compares features with their ratios of variance between classes and variance within classes	$\mathrm{Fisher}(k)=\frac{R_B^{(k)}}{R_w^{(k)}}$      $LS(f_i)=\frac{{\displaystyle \sum _{ab}{(f_{ra}-f_{rb})}^2{W}_{ij}}}{Var(f_r)}$
	ReliefF	Compares features with correlation between features and classes computed from ability of features to distinguish between close samples	$ReliefF(f_i,R1,R2)=\frac{|R1(A)-R2(A)|}{\max (A)-\min (A)}$
STF	FS	Obtains feature score with ability to distinguish positive classes and negative classes computed by average of both classes	$FS(i)=\frac{{({{\overline{f}}_i}^{(+)}-{\overline{f}}_i)}^2+{({{\overline{f}}_i}^{(-)}-{\overline{f}}_i)}^2}{\frac{1}{n_{+}-1}{\displaystyle \sum _{k=1}^{n_{+}}{({f_{k,i}}^{(+)}-{{\overline{f}}_i}^{(+)})}^2}+\frac{1}{n_{-}-1}{\displaystyle \sum _{k=1}^{n_{-}}{({f_{k,i}}^{(-)}-{{\overline{f}}_i}^{(-)})}^2}}$
	TS	Computes feature score with average and variance of features	$TS(i)=\frac{({{\overline{f}}_i}^{(+)}-{{\overline{f}}_i}^{(-)})}{\frac{1}{n_{+}-1}{\displaystyle \sum _{k=1}^{n_{+}}{({f_{k,i}}^{(+)}-{{\overline{f}}_i}^{(+)})}^2}+\frac{1}{n_{-}-1}{\displaystyle \sum _{k=1}^{n_{-}}{({f_{k,i}}^{(-)}-{{\overline{f}}_i}^{(-)})}^2}}$
SSL	MCFS	Combines cluster with feature coefficients of combinatorial classes to compute feature score	$MCFS(i)=\frac{\max }{k}\left|f_{k,i}\right|$
	Alpha	Evaluates features by dynamically adjusting threshold on error reduction to obtain selection results	$E(N_i)/E(M_i)<{\alpha }_{\Delta }/(1-{\alpha }_{\Delta })$
	Lasso	Uses L1 regularization to make weight of some learned features equal 0, to achieve purpose of sparse and feature selection	$Lasso(\stackrel{\wedge }{\beta })=arg min\left\{{\displaystyle \sum _{i=1}^n{\left(y_{\mathrm{i}}-{\beta }_0-{\displaystyle \sum _{j=1}^p{\beta }_j{x}_{ij}^{*}}\right)}^2+\lambda {\displaystyle \sum _{j=0}^p\left|{\beta }_j\right|}}\right\}$