![[2501.00773] Revisiting Graph Neural Networks for Graph-level Tasks: Taxonomy, Empirical Study, and Future Directions](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2501.00773] Revisiting Graph Neural Networks for Graph-level Tasks: Taxonomy, Empirical Study, and Future Directions
Summary
This article presents a comprehensive study on Graph Neural Networks (GNNs) for graph-level tasks, categorizing them into five types and proposing a unified evaluation framework, OpenGLT, to standardize assessments across diverse datasets and tasks.
Why It Matters
As GNNs are increasingly applied in various domains, understanding their strengths and weaknesses through a standardized evaluation framework is crucial for advancing research and practical applications. This study addresses the limitations of current evaluations, enhancing the reliability of GNNs in real-world scenarios.
Key Takeaways
- GNNs are categorized into five types for better understanding.
- OpenGLT framework standardizes evaluations across diverse graph tasks.
- The study conducts extensive experiments on 16 baseline models.
- Insights reveal strengths and weaknesses of existing GNN architectures.
- The findings aim to improve GNN applications in real-world scenarios.
Computer Science > Machine Learning arXiv:2501.00773 (cs) [Submitted on 1 Jan 2025 (v1), last revised 22 Feb 2026 (this version, v2)] Title:Revisiting Graph Neural Networks for Graph-level Tasks: Taxonomy, Empirical Study, and Future Directions Authors:Haoyang Li, Yuming Xu, Alexander Zhou, Yongqi Zhang View a PDF of the paper titled Revisiting Graph Neural Networks for Graph-level Tasks: Taxonomy, Empirical Study, and Future Directions, by Haoyang Li and 3 other authors View PDF HTML (experimental) Abstract:Graphs are fundamental data structures for modeling complex interactions in domains such as social networks, molecular structures, and biological systems. Graph-level tasks, which involve predicting properties or labels for entire graphs, are crucial for applications like molecular property prediction and subgraph counting. While Graph Neural Networks (GNNs) have shown significant promise for these tasks, their evaluations are often limited by narrow datasets, task coverage, and inconsistent experimental setups, hindering their generalizability. In this paper, we present a comprehensive experimental study of GNNs on graph-level tasks, systematically categorizing them into five types: node-based, hierarchical pooling-based, subgraph-based, graph learning-based, and self-supervised learning-based GNNs. To address these challenges, we propose a unified evaluation framework OpenGLT for graph-level GNNs. OpenGLT standardizes the evaluation process across diverse datasets, m...