This paper in ICLR describes a new attention mechanism for graph neural networks that builds off of the original multi-head attention for GATs. This mechanism is similar to part of the attention mechanism in Transformer networks that has shown great success. The authors incorporate edge information into query, key and value transformations to derive attention coefficients. Would be great if this could be in Pytorch Geometric.
rusty1s
commented
3 years ago
🚀 Feature
This paper in ICLR describes a new attention mechanism for graph neural networks that builds off of the original multi-head attention for GATs. This mechanism is similar to part of the attention mechanism in Transformer networks that has shown great success. The authors incorporate edge information into query, key and value transformations to derive attention coefficients. Would be great if this could be in Pytorch Geometric.