search

chnsh / DCRNN_PyTorch

Diffusion Convolutional Recurrent Neural Network Implementation in PyTorch
MIT License
454 stars 114 forks source link

The formulation of Diffusion Convolution is wrong #14

Closed hazdzz closed 3 years ago

hazdzz commented 3 years ago

In DCRNN, there is no diffusion process in the graph, because we cannot find any restart probability alpha in the code. Here, I offer a correct Diffusion Matrix as below:

GDC(PPR)

yuqirose commented 3 years ago

The probabilities are absorbed into the trainable weights in the neural networks.