Efficient Ensembles of Graph Neural Networks
TimeTuesday, July 12th1:30pm - 1:53pm PDT
Location3000, Level 3
Event Type
Research Manuscript
ML Algorithms and Applications
DescriptionEnsembles improve the accuracy and robustness of Graph Neural Networks (GNN), but suffer from high latency and storage requirements. We propose efficient GNN ensembles through Error Node Isolation (ENI). ENI identifies nodes that are likely to be incorrectly classified and suppresses their outgoing messages, leading to both computational and accuracy improvements. We propose techniques to create diverse ensembles, and demonstrate that ENI enables aggressive approximations of the models in the ensemble while maintaining accuracy. Our models are simultaneously up to 4.6% (3.8%) more accurate and up to 2.8X (5.7X) faster compared to non-ensemble (conventional ensemble) models, respectively.