Algorithm 2: Capsule Dual Graph CNN (C-DGCNN).

Input: FM, A, PPMI, yL, r, λ(t) and hidden convolution layers (H)

Output: Training model with best features.

Step 1: for t in range (0, Epoch number) do

Step 2:   $Z^A=ReLU\left(Con{v}_{LC}\right)$ using Eq. (16)

Step 3:   $Z^P=ReLU\left(Con{v}_{GC}\right)$ using Eq. (22)

Step 4:  Compute Loss using Eq. (23)

Step 5:  if convergence then break loops

Step 6:  end if

Step 7: end for

Step 8: Dynamic routing procedure (OCj|i,W′,r). // Where $W^{\mathrm{\prime }}\in {\mathbb{R}}^{m\times n}$ is the matrix of wights.

Step 9: $\overline{W^{\mathrm{\prime }}}\leftarrow W^{\mathrm{\prime }}+rand\left(size\left(W^{\mathrm{\prime }}\right)\right)$

Step 10: for all input capsules i and all output capsules j do

Step 11:   $O{C}_{j|i}\leftarrow \overline{{W^{\mathrm{\prime }}}_j}$ squash (OCj|i)) // Where $\overline{{W^{\mathrm{\prime }}}_j}\in {\mathbb{R}}^m$.

Step 12: end for

Step 13: for all input capsules i and all output capsules j do

Step 14:   $b_{ij}\leftarrow 0$

Step 15: end for

Step 16: for r iterations do

Step 17:  for all input capsules i do $c_i\leftarrow softmax(b_i)$

Step 18:  for all output capsules j do $s_j\leftarrow {\sum }_i{c}_{i,j}O{C}_{j|i}$

Step 19:  for all output capsule networks i do $O_i\leftarrow squash(s_i)$ // Where $squash\left(s_i\right)={\displaystyle \frac{{\Vert s_i\Vert }^2}{1+{\Vert s_i\Vert }^2}\cdot {\displaystyle \frac{s_i}{\Vert s_i\Vert }}}$.

Step 20:  for all input capsules i and output capsules j do

Step 21:   $b_{ij}\leftarrow b_{ij}+O{C}_{j|i}\cdot O_j$

Step 22: end for

Step 23: Return Oj

Step 24: end for