What the study found
LPS-GNN is a graph neural network framework designed to work on very large graphs. The abstract says it can perform representation learning on 100-billion-edge graphs with a single GPU in 10 hours and reports performance improvements in User Acquisition scenarios.
Why the authors say this matters
The authors suggest the framework matters because scalable graph neural networks often face a tradeoff between efficiency and prediction accuracy, especially on large graphs with neighbor explosion, which means a rapid growth in the number of neighboring nodes to process. They conclude that LPS-GNN offers a low-cost, flexible, and efficient approach.
What the researchers tested
The researchers examined existing graph partitioning methods and designed a new partition algorithm called LPMetis. They also proposed a subgraph augmentation strategy to improve predictive performance, and they tested the framework on public and real-world datasets as well as on Tencent's platform.
What worked and what didn't
According to the abstract, LPMetis outperformed current state-of-the-art approaches on various evaluation metrics. The framework is described as compatible with various GNN algorithms, and the paper reports performance lifts of 8.24% to 13.89% over state-of-the-art models in online applications, including a 13.8% improvement in User Acquisition scenarios.
What to keep in mind
The abstract does not provide detailed experimental limitations or failure cases. It also does not specify which exact graph neural network algorithms were used in every test, beyond saying the framework can accommodate various GNN algorithms.
Key points
- LPS-GNN is presented as a scalable graph neural network framework for very large graphs.
- The abstract says it can run representation learning on 100-billion-edge graphs with a single GPU in 10 hours.
- LPMetis, the new graph partition algorithm, is reported to outperform current state-of-the-art methods on evaluation metrics.
- The paper reports online application gains of 8.24% to 13.89% over state-of-the-art models.
- The abstract says the framework was deployed on Tencent's platform and tested on public and real-world datasets.
Disclosure
- Research title:
- LPS-GNN scales graph learning to 100-billion-edge graphs
- Publication date:
- 2026-03-31
- DOI:
- 10.1145/3801100
- OpenAlex record:
- View
- Image credit:
- Pexels • Google DeepMind · Pexels License
Get the weekly research newsletter
Stay current with peer-reviewed research without reading academic papers — one filtered digest, every Friday.