Can Classic GNNs Be Strong Baselines for Graph-level Tasks? Simple Architectures Meet Excellence

TL;DR

This paper reassesses graph-level modeling by enriching classic GNNs with six established techniques, forming the GNN$^+$ framework. Through systematic evaluation across 14 benchmark datasets, GCN$^+$, GIN$^+$, and GatedGCN$^+$ consistently reach top-three performances and often outperform state-of-the-art Graph Transformers while offering substantial efficiency advantages. Ablation analyses reveal that edge features, normalization, dropout, residual connections, FFN, and positional encoding each contribute meaningfully, with their importance varying by dataset scale and domain. Overall, the work challenges the assumption that complex GT architectures are necessary for graph-level excellence and highlights the practical potential of well-tuned, simple GNNs.

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.

Figures (6)

• Figure 1: Differences between graph-level and node-level tasks.
• Figure 2: The architecture of GNN$^{+}$.
• Figure 3: Experiments on the impact of the number of layers in GNN$^+$ on PATTERN, CLUSTER, and PascalVOC-SP. Top row: performance with residual connections; bottom row: performance without residual connections.
• Figure 4: Sensitivity analysis of dropout rates in GNN$^+$ on CLUSTER, PascalVOC-SP, Peptides-struct, and Peptides-func.
• Figure 5: t-SNE visualization of graph-level embeddings learned by GCN (left) and GCN$^+$ (right). Colors denote different classes.