Abstract
Convolutional graph networks are used in particle physics for effective eventreconstructions and classifications. However, their performances can be limitedby the considerable amount of sensors used in modern particle detectors ifapplied to sensor-level data. We present a pooling scheme that usespartitioning to create pooling kernels on graphs, similar to pooling on images.Partition pooling can be used to adopt successful image recognitionarchitectures for graph neural network applications in particle physics. Thereduced computational resources allow for deeper networks and more extensivehyperparameter optimizations. To show its applicability, we construct aconvolutional graph network with partition pooling that reconstructs simulatedinteraction vertices for an idealized neutrino detector. The pooling networkyields improved performance and is less susceptible to overfitting than asimilar network without pooling. The lower resource requirements allow theconstruction of a deeper network with further improved performance.