Results for multi-relational link prediction

[zhanqiuzhang]

Thanks for sharing the code!

Do you evaluate the performance of KE-GCN on the link prediction task? As far as I know, link prediction is the standard evaluation task for KGE models and many baseline models (e.g., R-GCN, W-GCN, KBGAT, CompGCN) in Table 1.

Looking forward to your reply.

[PlusRoss]

We did not evaluate our model on the link prediction task. The main reason is that according to previous works, GCN did not bring significant superior performance compared to non-GCN models such as RotatE and QuatE. The KBGAT model seems to have a much better performance, but recent work shows that their code has test data leakage problem. After fixing that, it failed to outperform RotatE and QuatE.