TopGQ: Post-Training Quantization for GNNs via Topology Based Node Grouping

Dain Kwon, Kanghyun Choi, Hyeyoon Lee, SunJong Park, Sukjin Kim, Jinho Lee

Keywords: Graph Neural Networks, Neural Network Quantization

TL;DR: We propose the first PTQ framework for graph neural networks, providing up to 358x less quantization time compared to the existing QAT baselines.

Abstract: Graph neural networks (GNN) suffer from large computational and memory costs in processing large graph data on resource-constrained devices. One effective solution to reduce costs is neural network quantization, replacing complex high-bit operations with efficient low-bit operations. However, to recover from the error induced by lower precision, existing methods require extensive computational costs for retraining. In this circumstance, we propose TopGQ, the first post-training quantization (PTQ) for GNNs, enabling an order of magnitude faster quantization without backpropagation. We analyze the feature magnitude of vertices and observe that it is correlated to the topology regarding their neighboring vertices. From these findings, TopGQ proposes to group vertices with similar topology information of inward degree and localized Wiener index to share quantization parameters within the group. Then, TopGQ absorbs the group-wise scale into the adjacency matrix for efficient inference by enabling quantized matrix multiplication of node-wise quantized features. The results show that TopGQ outperforms SOTA GNN quantization methods in performance with a significantly faster quantization speed.

Primary Area: learning on graphs and other geometries & topologies

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 4295