LPS-GNN : Deploying Graph Neural Networks on Graphs with 100-Billion Edges

Key findings from this study

• The study found that LPS-GNN processes 100 billion edge graphs in 10 hours using a single GPU while improving prediction accuracy in user acquisition scenarios.
• The authors report that the proposed LPMetis graph partitioning algorithm outperforms existing state-of-the-art methods on multiple evaluation metrics.
• The researchers demonstrate that the framework achieves 8.24% to 13.89% performance improvements over state-of-the-art models when deployed on real-world datasets.

Overview

LPS-GNN is a scalable graph neural network framework designed to perform representation learning on graphs containing up to 100 billion edges using a single GPU. The framework addresses computational bottlenecks inherent in iterative message-passing techniques and neighbor explosion issues in large-scale graphs.

Methods and approach

The authors examine existing graph partitioning methods and propose LPMetis, a superior graph partition algorithm optimized for large-scale processing. The framework incorporates a subgraph augmentation strategy to enhance predictive performance. LPS-GNN demonstrates compatibility with various GNN algorithms and has been deployed on the Tencent platform for evaluation.

Results

LPS-GNN processed 100 billion edge graphs in 10 hours on a single GPU, achieving a 13.8% performance improvement in user acquisition scenarios. Testing on public and real-world datasets showed performance lifts ranging from 8.24% to 13.89% over state-of-the-art models in online applications. LPMetis outperformed current partitioning approaches across multiple evaluation metrics.

Implications

The framework's ability to handle 100 billion-edge graphs on commodity hardware reduces infrastructure requirements for industrial-scale graph mining. The design's compatibility with multiple GNN algorithms enables broader adoption across diverse graph mining tasks beyond the tested user acquisition domain. Successful production deployment on the Tencent platform validates the framework's practical applicability in large-scale commercial systems.

Scope and limitations

This summary is based on the study abstract and available metadata. It does not include a full analysis of the complete paper, supplementary materials, or underlying datasets unless explicitly stated. Findings should be interpreted in the context of the original publication.