I would like to have a trainable weight matrix for each edge type in my graph and use this matrix in my
ScorePredictor class, similiar as in this DGL Tutorial 6.3.
I would like to multiply my source node feature z_i with the weight matrix M_r for that edge type and the target node feature z_j as in the following formula:
I have been looking through your DGL Message Functions but I haven’t found one which I could use for my idea. Since I do not have a feature vector for every edge, but a matrix for an edge type, the nomal u_mul_e, u_dot_e functions do not really work.
Is there any way I can use the built-in functions?
Thank you for any help in advance!