focalinterest.net

AI Summary of Scholarly Research

This page presents an AI-generated summary of a published research paper. The original authors did not write or review this article. [See full disclosure ↓]

LPS-GNN scales graph neural network training to 100-billion-edge graphs

Written by

editor@focalinterest.com

in

Research area:computer-science-aiai-ml

What the study found

The study introduces LPS-GNN, a graph neural network framework that is described as scalable, low-cost, flexible, and efficient. The abstract says it can perform representation learning on a 100-billion-edge graph with a single GPU in 10 hours and improve User Acquisition performance.

Why the authors say this matters

The authors say existing scalable graph neural network solutions struggle to balance execution efficiency and prediction accuracy because message passing requires heavy computation and large GPU memory, especially on graphs with many neighbors. They present LPS-GNN as a framework meant to address these challenges.

What the researchers tested

The researchers examined existing graph partitioning methods and designed a partition algorithm called Label Propagation with METIS, or LPMetis. They also proposed a subgraph augmentation strategy and built a framework that can accommodate various graph neural network algorithms. The article says the framework was tested on public and real-world datasets and deployed on the Tencent platform.

What worked and what didn't

LPMetis is reported to outperform current state-of-the-art approaches on various evaluation metrics. The abstract says the subgraph augmentation strategy improves predictive performance, and the overall framework achieved performance lifts of 8.24% to 13.89% over state-of-the-art models in online applications. The abstract does not describe any failed approach in detail.

What to keep in mind

The summary does not provide detailed limitations, and it does not specify the exact datasets or all evaluation settings in the abstract. The reported results are based on the abstract's description of public and real-world tests.

Key points

  • LPS-GNN is presented as a scalable graph neural network framework.
  • The abstract says it can run on a 100-billion-edge graph with a single GPU in 10 hours.
  • The authors report a 13.8% improvement in User Acquisition scenarios.
  • LPMetis is described as outperforming current state-of-the-art partitioning methods.
  • Online applications reportedly saw performance lifts of 8.24% to 13.89% over state-of-the-art models.

Disclosure

Research title:
LPS-GNN scales graph neural network training to 100-billion-edge graphs
Authors:
Xu Cheng, Liang Yao, Feng He, Yukuo Cen, Yufei He, Wenzheng Feng, Chenhui Zhang, Hongyun Cai, Jie Tang
Institutions:
American Institute for Economic Research, National University of Singapore, Sun Yat-sen University, Tencent (China), Tsinghua University, Tsinghua University, Tsinghua University, Zhipu AI (China), Zhipu AI (China)
Publication date:
2026-03-31
DOI:
10.1145/3801100
OpenAlex record:
View
AI provenance: This post was generated by gpt-5.4-mini (OpenAI). The original authors did not write or review this post.

More posts