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, insufficient architecture coverage, restricted task scope and scenarios, and inconsistent experimental setups, making it difficult to draw reliable conclusions across domains. 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. We propose a unified evaluation framework OpenGLT, which standardizes evaluation across four domains (social networks, biology, chemistry, and motif counting), two task types (classification and regression), and three real-world scenarios (clean, noisy, imbalanced, and few-shot graphs). Extensive experiments on 20 models across 26 classification and regression datasets reveal that: (i) no single architecture dominates both effectiveness and efficiency universally, i.e., subgraph-based GNNs excel in expressiveness, graph learning-based and SSL-based methods in robustness, and node-based and pooling-based models in efficiency; and (ii) specific graph topological features such as density and centrality can partially guide the selection of suitable GNN architectures for different graph characteristics.

翻译：暂无翻译