WWW 2022最佳论文:GNN的结构搜索系统



系统链接:GitHub - PKU-DAIR/SGL: A scalable graph learning toolkit for extremely large graph datasets. (WWW'22, 🏆 Best Student Paper Award)

This paper proposes PaSca, a new paradigm and system that offers a principled approach to systemically construct and explore the design space for scalable GNNs

本文提出了一种新的范式——PaSca 为可扩展的gnn提供系统构建和探索设计空间的原则性方法

we implement an auto-search engine that can automatically search well-performing and scalable GNN architectures



大多数的GNN需要perform a recursive neighborhood expansion to gather neural messages repeatedly,这个过程会造成指数增长的资源消耗,是目前大规模GNN计算面临的主要问题之一。

To demonstrate this issue, we utilize distributed training functions provided by DGL [2] to execute the train pipeline of GraphSAGE



  • Scalable Paradigm
  • Auto-search Engines.:

Concretely, the representatives (i.e.,PaSca-V2 and PaSca-V3) outperform the state-of-the-art JK-Net by0.2% and 0.4% in predictive accuracy on our industry dataset, while achieving up to 56.6× and 28.3× training speedups, respectively.

  • Relevance to Web.



  • GNN Pipelines.
  • Scalable GNN Instances.
  • Graph Neural Architecture Search.

提出了一个SGAP(Scalable Graph Neural Architecture Paradigm)范式:

it differs from the previous NMP and DNMP framework in terms of message type, message scale, and pipeline:

  • Pre-processing.
  • Model-training.
  • Post-processing.

Search Engine:


Evaluation Engine:

  • Graph Data Aggregator.
  • Neural Architecture Trainer.