Comparison with an alternative benchmark design

[skojaku]

This is a comment from @rachithaiyappa

wow did not know this paper existed. Looks like they did identify uniform negative edge sampling to be unfair! they show differences in common neighbors between positive edges and negative edges (https://openreview.net/attachment?id=YdjWXrdOTh&name=supplementary_material)

Since this paper already exists, perhaps we can show that these differences (of common neighbors) vanish in our degree biased sampling of negative edges setting?

I suggest doing this^ because it may otherwise invite a question as to how our new benchmark compares to the Heuristic Related Sampling Technique (HeaRT) benchmark which the references paper proposes

It makes sense. But do you think it is doable in one week?

[rachithaiyappa]

I dont think it is doable in one week

But perhaps just producing a plot like Fig4 here may be useful in case we are asked since this just requires just to calculate common neighbors (when negative edges are samples from biased sampling versus when they are sampled uniformly) and not rerun algorithms

If biased sampling better reduces the differences in common neighbors b/w positive and negative edges as compares to HEART, then amazing!

[skojaku]

Oh, I see. @MunjungKim show that the dependency on the transitivity decreases for our benchmark, so I think that our benchmark reduces the difference.

[rachithaiyappa]

awesome will be great to show that explicitly in supplementary and just write a short paragraph about why we did that analysis :) But ofcourse, I dont feel strongly about doing this but just a thought which crossed

[skojaku]

I think it is an important point. I only have 12 mins to work today so will revisit this point tomorrow morning. Please feel free to chime in everyone.