GAT with Initial Residual Connection and Random Walk with Restart

This is the implementation of the GAT with Initial Residual Connection and Random Walk with Restart (GAT-IRC-RWR) model.

Datasets

The datasets used in the experiments are available at the following links:

• Cora
• Citeseer
• PubMed

Code References

• GAT
• pytorch-GAT

Experiment Results

Cora Dataset:

Model	2 Layers	3 Layers	4 Layers	5 Layers	6 Layers	7 Layers	8 Layers
GAT	0.786	0.802	0.800	0.796	0.767	0.627	0.354
GATRWR	0.797	0.804	0.811	0.807	0.813	0.811	0.811
GATIRC	0.757	0.810	0.801	0.798	0.412	0.381	0.466
GATIRCRWR	0.747	0.802	0.808	0.803	0.807	0.809	0.809

Citeseer Dataset:

Model	2 Layers	3 Layers	4 Layers	5 Layers	6 Layers	7 Layers	8 Layers
GAT	0.688	0.684	0.650	0.624	0.629	0.554	0.371
GATRWR	0.667	0.703	0.700	0.696	0.698	0.699	0.703
GATIRC	0.687	0.688	0.681	0.612	0.378	0.422	0.380
GATIRCRWR	0.680	0.704	0.706	0.702	0.711	0.711	0.707

Pubmed Dataset:

Model	2 Layers	3 Layers	4 Layers	5 Layers	6 Layers	7 Layers	8 Layers
GAT	0.781	0.781	0.778	0.780	0.774	0.771	0.768
GATRWR	0.772	0.781	0.778	0.781	0.784	0.783	0.785
GATIRC	0.786	0.769	0.780	0.772	0.744	0.719	0.685
GATIRCRWR	0.779	0.769	0.774	0.777	0.781	0.783	0.777