Algorithm 1.

The training process of MDGCN-SRCNN.

Input : A labeled training data set$\left\{\text{X},\text{Y}\right\}={\left\{x_i,y_i\right\}}_{i=1}^N$, the maximum number of training epochs T; the initialize adjacency matrix A,regularization coefficientα.

Output: The learned adjacency matrix$\widehat{\text{A}}$, the model parameterΘ for MDGCN-SRCNN and the predicted labelŷ.

Step 1 : Initialize the model parametersΘ in MDGCN-SRCNN model. Set iteration unit iter = 1;

Step 2 : whileiter < T do

Step 3 : fork = 1, ..., l do

Step 4 : Calculate the k-th graph convolutional layer H(k)via Eq. (1) and calculate the k-th sum pooling layerPool(H(k));

Step 5 : fork = 1, ..., l do

Step 6 : Calculate the k-th SMR-based convolution layerCkvia Eq. (9);

Step 7 : Concatenate the different layers of features Fvia Eq. (10);

Step 8 : Calculate the prediction labelŷ via Eq. (11);

Step 9 : Update the adjacency matrix Aand the model parametersΘ via optimizer according to the cross-entropy loss.

Step 10: iter =iter+1;

Step 11: end while