Algorithm 1. KG-GAN model training for data imputation.

Input: Original complete traffic dataset ${\mathit{X}}_{\mathit{t}}(m\times n)$; road adjacency triplet $\mathit{R}$; attribute triplet $\mathit{A}TT$; attribute co-occurrence triplet $\mathit{R}\mathit{e}\mathit{l}\mathit{a}\mathit{t}\mathit{i}\mathit{o}\mathit{n}\mathit{s}$; loss hyperparameters $\mathit{\alpha }$; masking matrix $\mathit{M}$; indicator matrix $\mathit{H}$; the number of epochs $\mathit{N}$ and the initialized parameters: generator ${\mathit{\theta }}_{{\mathit{G}}_{\mathit{T}}}$, generator ${\mathit{\theta }}_{{\mathit{G}}_{\mathit{S}}}$, and discriminator ${\mathit{\theta }}_{\mathit{D}}$

1: Construct a KG = {R, ATT, Relations} triplet.

2: By maximizing the $P(R,ATT|X_E)$ obtain the knowledge representation matrix $X_E$.

3: Integrating $X_t$ and $X_E$ : $X_t^{\mathrm{\prime }}=EM$- $Cell(X_t,X_E)$

4: $\mathbf{f}\mathbf{o}\mathbf{r}$ epoch=1,2,…,N

5:    (1) Discriminator optimization:

6:       Obtain discriminator loss LD via Eqs. (10–13)

7:       Back-propagate LD to update ${\theta }_D$

8:    (2) Generator optimization:

9:       Obtain the output of $G_T\hat{X}$ via Eqs. (10), (11)

10:        $L_{R1}=\frac{1}{n}{\sum }_{i=1}^n||(X_t^{\mathrm{\prime }}-\hat{X})\odot M|{|}_2^2$

11:        $L_{G_T}\leftarrow -D(\hat{X},H)+\alpha L_{R1}$

12:       Obtain the output of $G_S\overline{X}$ via Eq. (12)

13:        $L_{R2}=\frac{1}{n}{\sum }_{i=1}^n||(X_t^{\mathrm{\prime }}-\overline{X})\odot M|{|}_2^2$

14:        $L_{G_S}\leftarrow -D(\overline{X},H)+\alpha L_{R2}$

15:       Back $\mathrm{\_}$propagate $L_{G_T}$, $L_{G_S}$ to update ${\theta }_{G_T}$, ${\theta }_{G_S}$

16: $\mathbf{e}\mathbf{n}\mathbf{d}\mathbf{f}\mathbf{o}\mathbf{r}$

17: Impute the missing values:

18: Obtain imputed data Ximputed via Eqs. (10–12)

Output: Trained parameters ${\theta }_{G_T}$, ${\theta }_{G_S}$, and ${\theta }_D$; imputed data Ximputed