Designing an Embedded Feature Selection Algorithm for a Drowsiness Detector Model Based on Electroencephalogram Data

. 2023 Feb 7;23(4):1874. doi: 10.3390/s23041874

Algorithm 2 Proposed FS Algorithm
Input: original EEG feature set [1 × d]
$n e t w o r k \leftarrow$ initialize	▹ FPL to ones, rest randomly
$n \leftarrow$ 1	▹ pruning step counter
$e p o c h s \leftarrow$ 0	▹ epochs between two pruning steps
$e p o c h s_f i n a l \leftarrow$ 0	▹ all epochs during training
$p r e c \leftarrow 0, p s o v f t \leftarrow 0$
while (d > $d e s_f e a t_n u m$ ) OR ( $e p o c h s_f i n a l < m a x_e p o c h s_f i n a l$ ) do
if ( $p r e c \geq f i n a l_p r e c$ ) OR ( $e p o c h s \geq m a x_e p o c h s$ ) OR ( $p s o v f t \geq m a x_p s o v f t$ ) then	▹ Perform pruning on the FPL
$N_{d e l e t e d_w e i g h t s}$ $\leftarrow f l o o r (\frac{f \cdot d}{n})$
d← delete the $N_{d e l e t e d_w e i g h t s}$ weights in the FPL with the smallest magnitude
epochs ← 0
n = n + 1
$p r e c, e p o c h s, e p o c h s_f i n a l \leftarrow$ train()
$n e w s e t_p r e c$ , $n e w s e t_p s o v f t$ ← take the new k-sized feature subset and train the base network on it from scratch