Confusion about lin_l and lin_r methods in SAGEConv class

[YiheWang]

Hi, I am confused about lin_l and lin_r methods in SAGEConv class.

Which one is W1, and which one is W2 respectively?

Besides, I found that W1 and W2 are actually the same in the original paper if apply the mean aggregator. They concatenate x_i and the mean of x_j before dotting with W. I am wondering if you follow the paper's idea or do some modifications?

Thanks in advance for your help!

[rusty1s]

lin_l refers to W_2, transforming the source/left features (sorry for poor wording), and lin_r refers to W_1. Note that transformation + summation is identical to concatenation + transformation. The former is more memory-efficient, that's why we use it here.

[YiheWang]

Note that transformation + summation is identical to concatenation + transformation.

I understand now. Thanks for your help!