About node labeling

[braid123]

Hi, Dr. Zhang

I have a question about the paper "Link Prediction Based on Graph Neural Networks". In appendix B, the paper refers to that If we do not remove y, d(i, x) will be upper bounded by d(i, y) + d(x, y), obscuring the “true distance” between i and x. How is this assumption valid? Or how is this sentence reflected in the code? How do i understand this sentence in the code or in the formula?

Thanks.

[muhanzhang]

Hi! It is reflected in this section of the code. When we compute all nodes' distance to target node 0, we temporarily remove the other target node 1 from the subgraph, and vice versa. This gives more diverse and conditional distance information, which is verified to improve the performance. See Appendix D of the labeling trick paper.