search

DEEP-PolyU / FeatWalk_AAAI19

Large-Scale Heterogeneous Feature Embedding
15 stars 3 forks source link

Question about FeatWalk #1

Open Tomposon opened 5 years ago

Tomposon commented 5 years ago

Brilliant idea.

I want to compute similarity matrix S, but the computation is expensive for the large-scale network.

I want to know that if I obtain the walk sequence by FeatWalk, in order to avoid computing S(reduce computing complexity), can I use the co-occurrence pairs to estimate the similarity matrix S.

Thank you very much!

xhuang31 commented 5 years ago

Thanks for your interest.

Yes, you can. It has two issues. First, the manipulations would still be expensive since you need to have a n by n matrix, n is the number of nodes. You would need to create this n by n matrix, and add 1 to the corresponding positions based on the co-occurrence pairs. Second, the precision of the learned similarity matrix would not be high.

Thanks.

Tomposon commented 5 years ago

Thanks for your great analysis. Yeah, in this way, the time complexity will be O(n^2).