P-DROP: Poisson-Based Dropout for Graph Neural Networks
Hyunsik Yun
TL;DR
This work tackles the over-smoothing issue in Graph Neural Networks by introducing a Poisson-process-based node selection mechanism that assigns each node an independent clock with rate $\lambda_v$, enabling asynchronous and structure-aware updates. It explores two applications: using Poisson-based updates as a dropout-like regularizer and as a dynamic subgraph training strategy, both aimed at maintaining structural diversity during propagation. Across Cora, CiteSeer, and PubMed, the proposed SGNN approach yields competitive accuracy, often matching or surpassing DropEdge and DropNode, with pronounced gains in later training stages on larger graphs. The method offers practical implications for scalable, lossless learning on graphs by balancing localized updates with global structural cues, while also providing insights into computation-cost trade-offs and future extension to subgraph propagation settings.
Abstract
Over-smoothing remains a major challenge in Graph Neural Networks (GNNs), where repeated message passing causes node representations to converge and lose discriminative power. To address this, we propose a novel node selection strategy based on Poisson processes, introducing stochastic but structure-aware updates. Specifically, we equip each node with an independent Poisson clock, enabling asynchronous and localized updates that preserve structural diversity. We explore two applications of this strategy: as a replacement for dropout-based regularization and as a dynamic subgraph training scheme. Experimental results on standard benchmarks (Cora, Citeseer, Pubmed) demonstrate that our Poisson-based method yields competitive or improved accuracy compared to traditional Dropout, DropEdge, and DropNode approaches, particularly in later training stages.