Enhanced Soups for Graph Neural Networks

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Zuber, Joseph, Sarkar, Aishwarya, Jennings, Joseph, Jannesari, Ali
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866916652817317888
author Zuber, Joseph
Sarkar, Aishwarya
Jennings, Joseph
Jannesari, Ali
author_facet Zuber, Joseph
Sarkar, Aishwarya
Jennings, Joseph
Jannesari, Ali
contents Graph Neural Networks (GNN) have demonstrated state-of-the-art performance in numerous scientific and high-performance computing (HPC) applications. Recent work suggests that "souping" (combining) individually trained GNNs into a single model can improve performance without increasing compute and memory costs during inference. However, existing souping algorithms are often slow and memory-intensive, which limits their scalability. We introduce Learned Souping for GNNs, a gradient-descent-based souping strategy that substantially reduces time and memory overhead compared to existing methods. Our approach is evaluated across multiple Open Graph Benchmark (OGB) datasets and GNN architectures, achieving up to 1.2% accuracy improvement and 2.1X speedup. Additionally, we propose Partition Learned Souping, a novel partition-based variant of learned souping that significantly reduces memory usage. On the ogbn-products dataset with GraphSAGE, partition learned souping achieves a 24.5X speedup and a 76% memory reduction without compromising accuracy.
format Preprint
id arxiv_https___arxiv_org_abs_2503_11612
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Enhanced Soups for Graph Neural Networks
Zuber, Joseph
Sarkar, Aishwarya
Jennings, Joseph
Jannesari, Ali
Machine Learning
Graph Neural Networks (GNN) have demonstrated state-of-the-art performance in numerous scientific and high-performance computing (HPC) applications. Recent work suggests that "souping" (combining) individually trained GNNs into a single model can improve performance without increasing compute and memory costs during inference. However, existing souping algorithms are often slow and memory-intensive, which limits their scalability. We introduce Learned Souping for GNNs, a gradient-descent-based souping strategy that substantially reduces time and memory overhead compared to existing methods. Our approach is evaluated across multiple Open Graph Benchmark (OGB) datasets and GNN architectures, achieving up to 1.2% accuracy improvement and 2.1X speedup. Additionally, we propose Partition Learned Souping, a novel partition-based variant of learned souping that significantly reduces memory usage. On the ogbn-products dataset with GraphSAGE, partition learned souping achieves a 24.5X speedup and a 76% memory reduction without compromising accuracy.
title Enhanced Soups for Graph Neural Networks
topic Machine Learning
url https://arxiv.org/abs/2503.11612