[論文レビュー] Can Classic GNNs Be Strong Baselines for Graph-level Tasks? Simple Architectures Meet Excellence

Yuankai Luo, Lei Shi+1 more|ArXiv.org|Feb 13, 2025

IoT and Edge/Fog Computing被引用数 3

ひとこと要約

Tldr: The paper shows that classic GNNs, when enhanced with a GNN + framework (edge features, normalization, dropout, residuals, FFN, and positional encoding), can achieve top performance on graph-level tasks across multiple benchmarks, rivaling or surpassing state-of-the-art Graph Transformers while being more efficient.

ABSTRACT

Message-passing Graph Neural Networks (GNNs) are often criticized for their limited expressiveness, issues like over-smoothing and over-squashing, and challenges in capturing long-range dependencies. Conversely, Graph Transformers (GTs) are regarded as superior due to their employment of global attention mechanisms, which potentially mitigate these challenges. Literature frequently suggests that GTs outperform GNNs in graph-level tasks, especially for graph classification and regression on small molecular graphs. In this study, we explore the untapped potential of GNNs through an enhanced framework, GNN+, which integrates six widely used techniques: edge feature integration, normalization, dropout, residual connections, feed-forward networks, and positional encoding, to effectively tackle graph-level tasks. We conduct a systematic re-evaluation of three classic GNNs (GCN, GIN, and GatedGCN) enhanced by the GNN+ framework across 14 well-known graph-level datasets. Our results reveal that, contrary to prevailing beliefs, these classic GNNs consistently match or surpass the performance of GTs, securing top-three rankings across all datasets and achieving first place in eight. Furthermore, they demonstrate greater efficiency, running several times faster than GTs on many datasets. This highlights the potential of simple GNN architectures, challenging the notion that complex mechanisms in GTs are essential for superior graph-level performance. Our source code is available at https://github.com/LUOyk1999/GNNPlus.

研究の動機と目的

• Motivate and investigate whether classic GNNs can excel in graph-level tasks when enhanced with established training techniques.
• Systematically evaluate GCN, GIN, and GatedGCN under the GNN + framework across diverse graph-level datasets.
• Analyze the contribution of each technique (edge features, normalization, dropout, residuals, FFN, and positional encoding) via ablation studies.

提案手法

• Introduce GNN +, an enhanced framework that combines six techniques into classic GNNs.
• Incorporate edge features into message passing and add a trainable edge feature module.
• Apply normalization (BatchNorm) and dropout to stabilize training and reduce overfitting.
• Incorporate residual connections to enable deeper architectures (3–20 layers).
• Append a feed-forward network (FFN) at each layer to boost expressiveness.
• Add positional encoding (RWSE) to inject graph structural information before learning.

実験結果

リサーチクエスチョン

• RQ1 Can classic GNNs (GCN, GIN, GatedGCN) reach SOTA performance on graph-level tasks when augmented with GNN +?
• RQ2 What is the impact of each GNN + component (edge features, normalization, dropout, residuals, FFN, positional encoding) on performance?
• RQ3 Do enhanced classic GNNs achieve top rankings across multiple graph-level benchmarks (GNN Benchmark, LRBG, OGB) compared to Graph Transformers?
• RQ4 What is the relative efficiency of enhanced GNNs versus Graph Transformers on large-scale graph-level datasets?

主な発見

• Enhanced versions of GCN, GIN, and GatedGCN achieve top-three rankings on all 14 graph-level datasets and first place on eight datasets.
• On ZINC, PATTERN, and CLUSTER, GNN + shows substantial improvements; on MNIST and CIFAR, GatedGCN + achieves top results among baselines.
• Across LRBG datasets, GCN + and GatedGCN + demonstrate strong performance, with GatedGCN + achieving top results on MalNet-Tiny and strong gains on PascalVOC-SP and COCO-SP.
• On Open Graph Benchmark (OGB) datasets, GatedGCN + ranks among the top three, with first place on ogbg-ppa and competitive performance on ogbg-molhiv and ogbg-molpcba.
• Ablation studies indicate that removing edge features, normalization, dropout, residuals, FFN, or positional encoding generally degrades performance, highlighting their contributions.
• The results collectively suggest that classic GNNs, when tuned with GNN +, can rival or surpass SOTA Graph Transformers while offering improved efficiency.