Training method

[AmosDinh]

Hello,
I am just wondering what might be the deviation from you method. Out of curiousity I implemented the HGT.
I use 100k users/bots for test and the rest for train. As node features I use the degree count per relationship type. I get an F1 of 66% on the test set (opposed to your 39%). It is implemented such that at test time, the graph contains both train and test users/bots.
Why do my results deviate this much?
Maybe because I also use the sampling approach from the HGT paper?
E.g.
supervision nodes seen: 6151840
F1. 0.6638356468073745
Prec. 0.5769753714634642
Rec. 0.781483414691525

Thank you