PaSca: A Scalable Graph Neural Architecture Search System for Large‑Scale Graph Learning

At the 2022 Web Conference (WWW2022), the joint Peking University‑Tencent Lab paper "PaSca: a Graph Neural Architecture Search System under the Scalable Paradigm" won the Best Student Paper Award, highlighting its novel contribution to scalable graph learning.

The authors identify two major bottlenecks of existing graph neural network (GNN) systems: limited scalability due to high communication costs in message‑passing mechanisms, and a high modeling barrier that requires expert knowledge to design network architectures for specific graph data and tasks.

To address these issues, the paper proposes a new Scalable Paradigm (SGAP) that separates message aggregation from model updates, defining a three‑stage workflow (pre‑processing → training → post‑processing). Based on SGAP, they construct a searchable design space containing over 150,000 distinct GNN architectures.

PaSca implements an automated search system composed of a search engine and a distributed validation engine. The search engine uses Bayesian optimization to recommend architectures that balance multiple objectives (prediction accuracy, memory usage, training/inference efficiency), while the validation engine efficiently evaluates candidates on distributed clusters.

Extensive experiments on ten public datasets demonstrate that (1) the SGAP‑based models achieve superior scalability compared with traditional message‑passing GNNs such as GraphSAGE, and (2) the architectures discovered by PaSca attain competitive or better predictive performance while meeting resource constraints.

The system has been deployed in Tencent’s internal Angel Graph platform on the Taiji machine‑learning infrastructure, powering applications such as video recommendation in WeChat articles, fraud detection, and personalized music recommendation, where it yields measurable improvements in click‑through rates and detection coverage.

In summary, PaSca introduces a novel GNN modeling paradigm, a large‑scale architecture search space, and an open‑source multi‑objective search framework (https://github.com/PKU-DAIR/SGL) that together advance the scalability and accessibility of graph neural networks for both research and industry.