GATv2
GATv2 is an improved version of the graph attention network that uses a dynamic attention mechanism to compute edge weights, with edge coefficients influenced by the graph structure.
Timeline (last 90 days)
The paper conducts experiments on nine datasets, with a model panel comprising GIN, GCN and GATv2 at depths ranging from 2 to 5 layers.
On the target logit, 10 of the 11 reportable GCN and GATv2 units show relative fidelity that depends on the evaluation intervention; in 9 of these units, each ranking achieves a larger deletion area under the intervention it was itself constructed for.
All Facts (2)
The paper conducts experiments on nine datasets, with a model panel comprising GIN, GCN and GATv2 at depths ranging from 2 to 5 layers.
50%UnverifiedOn the target logit, 10 of the 11 reportable GCN and GATv2 units show relative fidelity that depends on the evaluation intervention; in 9 of these units, each ranking achieves a larger deletion area under the intervention it was itself constructed for.
50%