Model Hyperparameters

For each model, we defaulted to using weighted cross entropy as our loss function and Adam as our optimizer, with a learning rate of 0.0001 and weight decay of 0.9. We also had a scheduler to lower the learning rate when loss plateaued for 15 epochs. Each model was run with 8 layers (except for SAN, which only had 4 due to memory constraints) and a batch size of 32 (Princeton was 8 due to memory constraints).

We ran our models on a mix of hardware, either a Nvidia A100 or a Nvidia RTX 3090.

Tabular Results

We used 5 different models, and 3 different optimizations on each of them (mixing and matching).

These tables display the test results of our models. Italics indicate the best train result (not listed here), and bold indicates the best test result for that dataset.

PascalVOC-SP (Macro F1 Score)

Modifier Type	GCN	GatedGCN	GIN	GAT	GAT (DW)	SAN
Baseline	0.0691	0.0885	0.1092	0.1462	0.0493	0.2006
Edge Adding	0.0510	0.0808	0.1058	0.1468	0.0493	0.1923
Laplacian PE	0.0787	0.0865	0.1112	0.1448	0.0493	0.2195
Random Walk SE	0.0799	0.0814	0.1124	0.1761	0.0493	0.2157
Laplacian PE + Edges	0.0764	0.0843	0.1143	0.1376	0.0493	0.2156
RWSE + Edges	0.0758	0.0850	0.1115	0.1792	0.0493	0.2105

Peptides-functional (Average Precision)

Modifier Type	GCN	GatedGCN	GIN	GAT	GAT (DW)	SAN
Baseline	0.4976	0.5166	0.5455	0.5245	0.4758	0.5745
Edge Adding	0.5214	0.5275	0.5540	0.5161	0.4609	0.5696
Laplacian PE	0.5326	0.5397	0.5643	0.5386	0.4784	0.6126
Random Walk SE	0.5211	0.5467	0.5689	0.5577	0.4860	0.5913
Laplacian PE + Edges	0.5133	0.5291	0.5634	0.5455	0.4957	0.6066
RWSE + Edges	0.5241	0.4701	0.5641	0.5395	0.4943	0.6092

Princeton Shape Benchmark (Average Precision)

Modifier Type	GCN	GatedGCN	GIN	GAT	GAT (DW)	SAN
Baseline	0.2228	0.1974	0.3440	0.2688	0.1649	0.3875
Edge Adding	0.1941	0.1429	0.2599	0.2123	0.1526	0.4270
Laplacian PE	0.3497	0.3843	0.3842	0.3472	0.2693	0.4268
Random Walk SE	0.3195	0.3921	0.3020	0.2951	0.2761	0.4415
Laplacian PE + Edges	0.3834	0.3089	0.3768	0.3382	0.3079	0.4493
RWSE + Edges	0.3327	0.3358	0.3932	0.3357	0.3003	0.4629