Algorithm A1: MGANet for sleep staging

Input: An information network: $G_A=(N,C,A)$

Output: Node vector fusion information representations ${\overrightarrow{f}}_{*}^{\prime }$

Initialization: $i=0$, training epochs I and labeled nodes $y_L$

While $i\le I$ do

for a vertex $v_n\in V$ do

${\overrightarrow{f}}_{s,n}^{}\leftarrow \sigma \left(\frac{1}{L}{\displaystyle \sum _{l=1}^L{\displaystyle \sum _{N_k\in m}{\alpha }_{N_c,N_k}^{(l)}{\mathrm{W}}^{(l)}{\overrightarrow{f}}_{N_k}}}\right)$//Learn sleep vector for each spatial feature.

${\overrightarrow{f}}_{s,n}^A\leftarrow (F_1,F_2,\cdots ,F_n)A_S^{\prime }$//Learn attention spatial feature representation.

${\overrightarrow{f}}_{t,n}^{}\leftarrow$//Learn the sleep vector for each temporal feature by Equation (8).

${\overrightarrow{f}}_{t,n}^A\leftarrow (F_1,F_2,\cdots ,F_{T_{r-1}})A_S^{\prime }$//Learn attention temporal feature representation.

end

$u\leftarrow a({\mathrm{W}}_{N_c}^G{\overrightarrow{f}}_{N_c},{\mathrm{W}}_{N_k}^G{\overrightarrow{f}}_{N_k})$//Calculate attention coefficient.

$P_{{\mathrm{f}}^{\prime }}\leftarrow \mathrm{softmax}(F_C(G({\overrightarrow{f}}^{\prime })))$//sleep stage probability distribution.

${\overrightarrow{f}}_{*}^{\prime }\leftarrow A_{{\mathrm{f}}^{\prime }}\otimes P_{{\mathrm{f}}^{\prime }}$//epoch features and transitional fusion information.

$i=i+1$

end