Algorithm 1.

implementation of OLB-CMI.

Input : Full feature set X = {X1, X2, X3, …, XD}, class label C, the number of features to be selected d and threshold α

index set of the selected features: S = {}

index set of the unselected features: S̄ = {1, 2, 3, …, D}

for i = 1 to D do

fea_lab_mi[i] = I(Xi, C)

fea_entropy[i] = H(Xi)

end for

$k^{\ast }=\underset{1\le i\le D}{\text{argmax}}(\text{fea}\_\text{lab}\_\text{mi}[i])$

S[1] = k*

S̄ = S̄/k*

for i = 2 to d do

for k = 1: D − i + 1 do

fea_lab_mi [i −1] [ S̄ [k)] = I(XS[i − 1], C; XS̄[k])

$\mathrm{t}=\underset{1\le m\le i-1}{\text{argmax}}(\text{fea}\_\text{lab}\_\text{mi}[\mathrm{m}][\overline{\mathit{S}}[k]])$

if $\left(\frac{I(X_{\mathit{S}[\mathrm{t}]},C;X_{\overline{\mathit{S}}[k]})}{\text{fea}\_\text{entropy}[\overline{\mathit{S}}[k]]}\ge \alpha \right)$ then

max_fea_lab_cmi[S̄[k]] = I(XS[t], C; XS̄[k]) − I(XS[t]; XS̄[k])

else

max_fea_lab_cmi [S̄[k]] = 0

end if

end for

$k^{\ast }=\underset{1\le k\le D-1}{\text{argmax}}(\text{max}\_\text{fea}\_\text{lab}\_\text{cmi}[\overline{\mathit{S}}[k]])$

S[i] = S̄[k*]

S̄ = S̄/S̄[k*]

end for

output: index set of the selected features S