Homophily Level

Homophily level, the tendency of similar nodes to connect in a network, is a crucial factor influencing the performance and fairness of graph neural networks (GNNs). Current research focuses on understanding how variations in local homophily, differing from the overall network homophily, impact GNN accuracy and fairness, particularly in the context of heterophilic (dissimilar node connections) graphs. Researchers are developing new GNN architectures and algorithms, including those incorporating random walk aggregation and global information aggregation, to address the challenges posed by varying homophily levels and improve robustness against adversarial attacks. This work has significant implications for various applications relying on GNNs, such as node classification and metagenomic binning, by improving model accuracy and mitigating biases.