This is the official repository for the paper Contextualized Messages Boost Graph Representations.
SIR-GCN emphasizes the non-linear and contextualized transformation of neighborhood features to control and guide hash (aggregation) collisions when the space of node feature is uncountable. The model may be expressed as
where
All experiments are conducted on a single Nvidia Quadro RTX 6000 (24GB) card using the Deep Graph Library (DGL) with PyTorch backend.