reproducing Disease LP visualization

[xinyue96]

Hi, I am trying to reproduce Figure 3 (b) in your paper, and what I did was first train Disease_lp data with the hyperboloid manifold for 3 dimension, python train.py --task lp --dataset disease_lp --model HGCN --lr 0.01 --dim 3 --num-layers 2 --num-layers 2 --act relu --bias 1 --dropout 0 --weight-decay 0 --manifold Hyperboloid --normalize-feats 0 --log-freq 5 And then use your function in the manifold module to convert it to the 2 dimension poincare embedding. But when I plot the 2d embedding, the points are distributed only in one quarter of the disk and does not look like figure 3 (b) at all. Could you please tell me if I am doing it the way you did it? Thanks!

[ines-chami]

Hi! Please try to remove the relu activation for visualization.

[xinyue96]

Hi! Please try to remove the relu activation for visualization.

Thank you, that makes sense. But then it is correct to say that for non-visualization purpose (LP or NC), HGCN is basically only using the positive quadrant of the hyperbolic space for training embedding because of the activation?

[ZhangKaly]

Hi, I am trying to reproduce Figure 3 (b) in your paper, and what I did was first train Disease_lp data with the hyperboloid manifold for 3 dimension, python train.py --task lp --dataset disease_lp --model HGCN --lr 0.01 --dim 3 --num-layers 2 --num-layers 2 --act relu --bias 1 --dropout 0 --weight-decay 0 --manifold Hyperboloid --normalize-feats 0 --log-freq 5 And then use your function in the manifold module to convert it to the 2 dimension poincare embedding. But when I plot the 2d embedding, the points are distributed only in one quarter of the disk and does not look like figure 3 (b) at all. Could you please tell me if I am doing it the way you did it? Thanks!

Hi Xinyue, would you be able to share your code for this reproduction? Any help would be really appreciated!