Skip to content

Latest commit

 

History

History
56 lines (36 loc) · 3.09 KB

Additional_experiment.md

File metadata and controls

56 lines (36 loc) · 3.09 KB

Edgeless-GNN-external

Here, we provide additional experimental results in "Edgeless-GNN :Unsupervised Inductive Edgeless Network Embedding".

Additional experimental results 1

Performance of Edgeless-SAGE on Cora with different number of layers.

AP AUC macro F1 micro F1 NMI
Single layer 0.8929 +/- 0.0140 0.8905 +/- 0.0127 0.6783 +/- 0.0335 0.7177 +/- 0.0343 0.5109 +/- 0.0212
Two layers 0.8464 +/- 0.0142 0.8590 +/- 0.0137 0.6254 +/- 0.0290 0.6665 +/- 0.0267 0.4408 +/- 0.0540
Three layers 0.7329 +/- 0.0221 0.7443 +/- 0.0310 0.4392 +/- 0.0549 0.5177 +/- 0.0422 0.3354 +/- 0.0415

Additional experimental results 2

Comparison of different architectures on Citeseer dataset.

Architecture AP AUC macro F1 micro F1 NMI
Edgeless-SAGE 0.9394 +/- 0.0006 0.9318 +/- 0.0072 0.5675 +/- 0.0378 0.6502 +/- 0.0299 0.4489 +/- 0.0506
Edgeless-GCN 0.8921 +/- 0.0131 0.8892 +/- 0.0112 0.2554 +/- 0.0240 0.4943 +/- 0.0423 0.2695 +/- 0.0590
Edgeless-GIN 0.8633 +/- 0.0146 0.8752 +/- 0.0124 0.5567 +/- 0.0370 0.6687 +/- 0.0362 0.2775 +/- 0.0395

Additional experimental results 3

Comparison of different choices of k on Citeseer dataset.

k AP AUC macro F1 micro F1 NMI
2 0.9376 +/- 0.0077 0.9279 +/- 0.0072 0.5752 +/- 0.0398 0.6545 +/- 0.0250 0.4184 +/- 0.0326
3 0.9385 +/- 0.0062 0.9313 +/- 0.0072 0.5698 +/- 0.0463 0.6511 +/- 0.0487 0.4126 +/- 0.0377
4 0.9325 +/- 0.0103 0.9282 +/- 0.0095 0.5589 +/- 0.0564 0.6672 +/- 0.0388 0.4336 +/- 0.0351
5 0.9365 +/- 0.0082 0.9284 +/- 0.0081 0.5508 +/- 0.0494 0.6620 +/- 0.0377 0.4589 +/- 0.0453
6 0.9385 +/- 0.0063 0.9304 +/- 0.0078 0.5767 +/- 0.0456 0.6507 +/- 0.0407 0.4253 +/- 0.0269

Additional experimental results 4

Effect of alpha and beta on Citeseer dataset.

Additional experiment 3

Additional experimental results 5

Comparison with [40] on node classification. We have used the author's implementation with modification to 1) Edge deletion mechanism (to generate edgeless nodes) 2) Train/val/test split to match our setting.

Dataset Method micro F1
Cora Edgeless-SAGE 0.7177 +/- 0.0343
LDS-GNN 0.2777 +/- 0.0693
Citeseer Edgeless-SAGE 0.6697 +/- 0.0299
LDS-GNN 0.4791 +/- 0.1367