How the column prop is calculated?

[swyu0711]

The column norm is squared in paper, while in code it is not squared as frobenious norm. It seems like a contradiction. I changed it to squared form as in the paper and found that accuracy decreased a lot.

[matenure]

Neither squaring nor nonsquaring guarantees variance reduction, as implied in the paper (see the paragraph above proposition 4). It’s up to empirical tuning to adopt the best sampling distribution. We did a quick check with cora and pubmed (using pubmed_Mix_sampleA.py) and found that the results between squaring and nonsquaring were quite comparable, with the latter slightly better. Thank you for spotting this!