Algorithm 1: TF-SCN

Input: Training dataset $D=\left\{\right(x_i,y_i\left)\right\}$; Lmax is the maximum number of nodes in the hidden layer across both branches; maximum candidate trials Tmax; residual tolerance $E=[e_1,e_2,\cdots ,e_C]$; tolerable error σ; regularization coefficient $\lambda$; random parameter bound η.

Output: Temporal parameters ΘT, frequency-sensitive parameters ΘF, output weight β, and prediction results.

1. Initialize HT = ∅, HF = ∅, ΘT = ∅, ΘF = ∅, LT = 1, LF = 1.

2. while (LT < Lmax or LF < Lmax) and ||E|| > σ do:

3.       Initialize candidate set Φ = ∅, Ω = ∅, t = 0;

4.       while (ΘT and ΘF are empty) and t < Tmax do:

5.          Generate a random number r on the interval [0, 1];

6.          Solve the sequence value: ${\mu }_L=(1-r)/(1+L_T+L_F)$;

7.          Generate $w_t^T$, $b_t^T$ randomly in [−η, η];

8.          Generate $w_t^F$, $b_t^F$ randomly in [−η, η];

9.          Construct temporal feature representation for the candidate node according to Equation (2);

10.           Construct frequency feature representation for the candidate node according to Equation (3);

11.           Calculate ${\epsilon }_q^T$ and ${\epsilon }_q^F$ according to Equation (8);

12.           if $\min \left({\epsilon }_1^T,{\epsilon }_2^T,\cdots ,{\epsilon }_C^T\right)>0$:

13.             Append $(w_t^T,b_t^T)$ to the set Φ;

14.             LT = LT + 1;

15.           end if

16.           if $\min \left({\epsilon }_1^F,{\epsilon }_2^F,\cdots ,{\epsilon }_C^F\right)>0$:

17.             Append $(w_t^F,b_t^F)$ to the set Ω;

18.             LF = LF + 1;

19.           end if

20.           t = t + 1

21.      end while

22.   Add $(w_t^T,b_t^T)$ and $(w_t^F,b_t^F)$ into ΘT, ΘF;

23.   Solve β according to Equation (7);

24.   Update residual matrix based on Equation (6);

25.end while