How to calculate the gradient of metapath embedding?

[sh0416]

If we use binary step function, the gradient of step function is always zero.

I think there is no backpropagation for the metapath embedding vector.

However, there is a update in metapath vector and I don't understand the flag "is_deepwalk".

What is that?

[sh0416]

To my knowledge, this implementation update the gradient of sigmoid function to metapath vector no matter what the regularization function we choose.

Is it intended?

[csiesheep]

Hi, Thanks for asking. Yes, I use the gradient of sigmoid function to approximate the gradient of step function.