Algorithm 2. Model training.

1:   Initialize model parameters randomly as $\mathsf{\Theta }$, and total loss as Loss = 0;

2:   Repeat

3:      For each ${(Q,Y}_{t,i})$ in ${\mathcal{D}}_{train}$

4:         Perform forward propagation to compute $\widehat{Y_{t,i}}$;

5:         Compute mean square error $L\left(\mathsf{\Theta }\right)={‖Y_{t,i}-\widehat{Y_{t,i}}‖}^2$;

6:         $Loss+=L\left(\mathsf{\Theta }\right)$;

7:         Perform backward propagation to compute$∆\mathsf{\Theta }$;

8:         update $\mathsf{\Theta }$, $\mathsf{\Theta }=\mathsf{\Theta }+∆\mathsf{\Theta }$;

9:      End

10:   Until $Loss<\sigma$;

11:   Output: Trained model DPVEM-DGD