Algorithm 1: Recursive Feature Selection

1: Step 1: Start with $m=1$, and find the best ${\mathit{S}}_m$ using Equation (5) from all feature subsets 2: $\mathit{S}{\mathit{S}}_m$. 3: Step 2: For $m\le d$, sort all feature subsets in $\mathit{S}{\mathit{S}}_m$ by the objective function in Equation (5). 4: Select only the first $min\left(l,\left(\begin{array}{c}72\\ m\end{array}\right)\right)$ feature subsets based on performance, where $l$ is the user-5: defined quota. In this study, $l$ is set to be 500. 6: Step 3: For every selected subset ${\mathit{S}}_m$ from Step 2, find the absolute complement of ${\mathit{S}}_m$, 7: represented by ${{\mathit{S}}^{\prime }}_m$. For each feature ${\mathit{f}}_i\in {{\mathit{S}}^{\prime }}_m$, create a new subset ${\mathit{S}}_{m+1}$ by adding ${\mathit{f}}_i$ 8: to ${\mathit{S}}_m$; and keep this new subset ${\mathit{S}}_{m+1}$ only if its performance is better than ${\mathit{S}}_m$. 9: Step 4: Insert all remaining ${\mathit{S}}_{m+1}$ from step 3 to new feature set $\mathit{S}{\mathit{S}}_{m+1}$. Increase m by 1. 10: Step 5: Repeat step 2 to 4 until $m=d$.