Post #445

GML Newsletter: Homophily, Heterophily, and Oversmoothing for GNNs Apparently, Cora and OGB datasets are mostly assortative datasets, i.e. nodes of the same labels tend to be connected. In many real-world applications, it's not the case, i.e. nodes of different groups are connected, while within the groups the connections are sparse. Such datasets are called disassortative graphs. What has been realized in 2020 and now in 2021 is that typical GNNs like GCN do not work well in disassortative graphs. So several GNN architectures were proposed to get good performance for these datasets. Not only these new GNNs work well on assortative and disassortative graphs, but also they solve the problem of oversmoothing, i.e. effectively designing many layers for GNNs. In my new email newsletter I discuss this change from assortative to disassortative GNNs and its relation to oversmoothing. What's interesting is that existing approaches still do not rely explicitly on the labels, but rather learn parameters to account for heterophily. In the future, I think there will be more hacks how to integrate target labels directly into the GNN algorithm.