P-DROP: Poisson-Based Dropout for Graph Neural Networks

ICLR 2026 Conference Submission13199 Authors

Keywords: Graph Neural Networks, Probability Theory, Interacting Particle Systems, Dropout, Poisson Process

TL;DR: Applying an idea from mathematical probability model "Interacting Particle Systems" to Graph Neural Networks.

Abstract:

Stochastic processes are widely used in machine learning, yet interacting particle systems—a class of stochastic processes—have seen limited application. In this paper, we leverage an idea from classical interacting particle systems to propose a novel node selection strategy based on Poisson processes. By equipping each node with an independent Poisson clock, our method enables asynchronous and localized updates that preserve structural diversity. This approach introduces not only stochastic but also structure-aware dynamics to graph training.

Recent work has introduced various drop-based techniques such as DropNode, DropEdge, and DropMessage to inject randomness and improve generalization in graph neural networks. Our Poisson-based method offers a principled alternative to these heuristics, yielding competitive or improved performance while grounding the stochasticity in a well-defined process. This work bridges probability theory and graph learning, opening a new avenue for principled stochastic design in GNNs.

Primary Area: learning on graphs and other geometries & topologies

Submission Number: 13199