Graph Aggregation Beyond Homophily Assumption - a more meaningful way to model networks

Lucy Lu

Keywords: Graph Neural Networks, Homophily Assumption, Node Aggrega- tion, Ricci Curvature

Abstract: Data aggregation on Homophily/Heterophily networks have caused lots of discussions. Existing solutions are all based on Homophily assumption that Heterophily edges are considered as noisy data and need to be eliminated. In this paper, we first conduct a case study to show data aggregation can not be affected by network types, but aggregation strategies. Graph Weighted Aggregation (GWA) method is proposed to perform aggregation with three attributes: (1) node features, (2) network topology and (3) label information. We also propose to use Riemannian manifold to model topological networks with Ricci Curvature as the force of influence between adjacent nodes. The three attributes together can formulate an ag- gregation strategy and update through message passing on Graph Neural Network until the network reaches the minimum energy. This methodology defines a more meaningful way to aggregate neighboring nodes with no regard to Homophily assumption. GWA algorithm outperforms the state-of-the-art algorithms on bench- mark datasets.

Submission Number: 20