[Paper Review] Graph Property Prediction on Open Graph Benchmark: A Winning Solution by Graph Neural Architecture Search

Xu Wang, Huan Zhao+2 more|arXiv (Cornell University)|Jul 13, 2022

Advanced Graph Neural Networks4 citations

TL;DR

This paper proposes a neural architecture search framework combining Pooling Architecture Search (PAS) and F²GNN to automatically discover optimal graph neural network architectures for graph property prediction. By jointly searching aggregation functions and GNN topology with learnable feature fusion, the method achieves state-of-the-art performance on three OGB benchmark datasets—ogbg-molhiv, ogbg-molpcba, and ogbg-ppa—surpassing fixed-architecture models and demonstrating strong generalization across molecular and biological graph tasks.

ABSTRACT

Aiming at two molecular graph datasets and one protein association subgraph dataset in OGB graph classification task, we design a graph neural network framework for graph classification task by introducing PAS(Pooling Architecture Search). At the same time, we improve it based on the GNN topology design method F2GNN to further design the feature selection and fusion strategies, so as to further improve the performance of the model in the graph property prediction task while overcoming the over smoothing problem of deep GNN training. Finally, a performance breakthrough is achieved on these three datasets, which is significantly better than other methods with fixed aggregate function. It is proved that the NAS method has high generalization ability for multiple tasks and the advantage of our method in processing graph property prediction tasks.

Motivation & Objective

• To address the challenge of designing optimal GNN architectures for graph property prediction across diverse graph types.
• To overcome the over-smoothing problem in deep GNNs through adaptive architecture search and feature fusion.
• To improve performance on OGB's graph-level classification benchmarks without relying on external datasets.
• To demonstrate the generalization capability of neural architecture search (NAS) across multiple graph learning tasks.

Proposed method

• Introduces a differentiable search framework for pooling operations and aggregation functions via Pooling Architecture Search (PAS).
• Enhances PAS with F²GNN, a topology search method that learns optimal GNN connectivity patterns through differentiable operations.
• Applies learnable feature selection and fusion mechanisms to preserve node representation discrimination in deep GNNs.
• Uses a search space that includes multiple GAT variants and residual-like connectivity patterns to improve expressivity.
• Employs a two-stage search: first, optimize aggregation functions per layer; second, search topology with F²GNN for deeper models.
• Leverages differentiable architecture search to jointly optimize message-passing functions and graph topology for improved inductive bias.

Experimental results

Research questions

• RQ1 Can neural architecture search (NAS) outperform fixed GNN architectures on graph property prediction tasks across diverse graph types?
• RQ2 How effective is the combination of PAS and F²GNN in mitigating over-smoothing in deep GNNs?
• RQ3 Does the proposed method generalize across molecular and biological graph datasets without task-specific architectural tuning?
• RQ4 Can learnable feature fusion and adaptive topology design improve performance on large-scale graph benchmarks like OGB?
• RQ5 What is the impact of architecture search on model performance when external data is not used?

Key findings

• The proposed method achieved 1st place on ogbg-molhiv with a test ROC-AUC of 0.8420 ± 0.0015, outperforming Graphormer and CIN.
• On ogbg-molpcba, the method achieved a test AP of 0.3147 ± 0.0015, significantly surpassing PAS (0.3012) and other SOTA models.
• For ogbg-ppa, the method reached 82.01% test accuracy, exceeding ExpC* (81.40%) and PAS (78.28%) by a substantial margin.
• The integration of F²GNN into PAS improved performance on all three datasets, confirming the value of topology search in deep GNNs.
• The method achieved SOTA results without using external datasets, proving its effectiveness and generalization in real-world graph learning.
• The ablation study confirms that both PAS and F²GNN contribute independently and synergistically to performance gains across tasks.