search

acbull / pyHGT

Code for "Heterogeneous Graph Transformer" (WWW'20), which is based on pytorch_geometric
MIT License
785 stars 162 forks source link

Question about parameters in Eq.2 in paper #53

Open leafskywe opened 1 year ago

leafskywe commented 1 year ago

Dear authors, I have been reading this paper these days, and I really admire your pioneering work in the field of heterogeneous graph. My question is: Does the function or effect of Wϕ(e) and µ<s,e,t> overlap? The explanation of Wϕ(e) in paper is "capture different semantic relations even between the same node type pairs", but in my understanding, µ<s,e,t> also serve this purpose. For example, given two meta relations<s,e1,t> and <s,e2,t>, Wϕ(e) can distinguish between e1 and e2 while µ can between distinguish <s,e1,t> and <s,e2,t>, but doesn't distinguishing between <s,e1,t> and <s,e2,t> mean that you also distinguish between e1 and e2 at same time? So I hope you can explain the difference between Wϕ(e) and µ according to my confusion. In addition, I would like to know whether you have done relevant experiments to verify that Wϕ(e) and µ are both indispensable? if not, I'd like to try to do this, such as just keeping one of them. I would appreciate your response!

acbull commented 1 year ago

Hi: µ in our implementation is a single scalar, not a matrix. So it could only learns global weight for a particular meta-relation, not a full transformation matrix.

leafskywe commented 1 year ago

Oh,I got it. Thanks a lot!