Homophily and Heterophily in Semi-supervised Node Classification: AMUD and ADPA
Konsep Inti
Developing a powerful GNN model that can ensure performance under both homophily and heterophily is crucial for efficient graph learning.
Abstrak
Graph neural networks (GNNs) have shown significant performance in semi-supervised node classification.
Existing GNNs struggle to achieve optimal node representations due to the constraints of undirected graphs.
AMUD quantifies the relationship between node profiles and topology, offering valuable insights for adaptively modeling directed graphs.
ADPA introduces a new directed graph learning paradigm, achieving effective message aggregation.
Extensive experiments on benchmark datasets substantiate the impressive performance of ADPA.
Breaking the Entanglement of Homophily and Heterophily in Semi-supervised Node Classification
"Undirected GNNs are more suitable for handling homophilous undirected graphs, while directed GNNs exhibit a significant advantage in dealing with heterophilous digraphs."
"Modeling directed information assists directed GNNs in capturing intricate heterophily while discarding directed information is more crucial for utilizing homophily."