Reimplementation of NeurIPS 2020 paper "Beyond homophily in graph neural networks: Current limitations and effective designs" based on PyTorch and PyTorch Geometric (PyG).
python main.py
- Besides the original idea of H2GCN, we also try another variant version of it by changing the original GCN-based weighted aggregator to a GCN layer with learnable parameters (labeled as H2GCN-Variant in this project). However, the original H2GCN performs much better than the H2GCN-Variant in this implementation.
- Currently, the number of aggregation hops and the number of layers are both simply set as 2. It may need necessary changes to our codes if the above two parameters change.