RADE: Random Add-Drop Edge as a Regularizer
Researchers have introduced RADE, a novel technique for Graph Neural Networks (GNNs) designed to combat overfitting and improve the handling of long-range information. Unlike previous methods that focus on either regularization or connectivity, RADE simultaneously drops and adds edges to the graph during training. This dual approach aims to align training and inference, preventing distribution shifts while enhancing communication across distant nodes. The method is further optimized with an adaptive gradient-norm balancing algorithm, making it practically hyperparameter-free and demonstrating strong performance on classification benchmarks. AI
IMPACT Introduces a novel regularization technique for GNNs that addresses both overfitting and long-range information processing.