PaSca: A Scalable Graph Neural Architecture Search System

Problem : Industrial applications demand timely and scalable processing of massive graph data, yet traditional message‑passing GNNs suffer from high communication and computation costs, leading to low scalability and high modeling barriers.

Method : PaSca implements an end‑to‑end searchable system that leverages the Scalable Graph Neural Architecture Paradigm (SGAP), which separates processing into pre‑processing, training, and post‑processing stages, drastically reducing communication rounds and allowing arbitrary downstream models (MLP, DNN, tree models). The system automatically searches for optimal GNN configurations under multiple objectives (accuracy, training time, memory) using Bayesian optimization.

Experiments : Extensive evaluations on standard benchmarks, larger datasets, and industry data demonstrate that SGAP‑based GNNs achieve near‑linear speed‑up and superior trade‑offs between accuracy and inference time compared to traditional message‑passing models such as GraphSAGE. Representative PaSca‑derived architectures (V2, V3) consistently outperform baselines in both scalability and prediction performance.

Conclusion : PaSca successfully automates the design of billion‑node GNNs, has been deployed on Tencent’s Taiji platform for recommendation and risk control, and its codebase (SGL) is open‑source. The framework not only achieves state‑of‑the‑art results on OGB leaderboards but also provides a flexible tool for researchers to explore scalable GNN designs.