Revisiting Graph Neural Networks for Graph-level Tasks: Taxonomy, Empirical Study, and Future Directions

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, multiple graph tasks (e.g., classification and regression), and real-world scenarios, including noisy, imbalanced, and few-shot graphs. Extensive experiments are conducted on 16 baseline models across five categories, evaluated on 13 graph classification and 13 graph regression datasets. These experiments provide comprehensive insights into the strengths and weaknesses of existing GNN architectures.

💡 Research Summary

The paper presents a thorough re‑examination of Graph Neural Networks (GNNs) for graph‑level tasks—tasks that require a single prediction for an entire graph, such as molecular property prediction, subgraph counting, or graph classification. The authors first identify a set of recurring shortcomings in existing benchmark studies: (1) the lack of a clear taxonomy that distinguishes GNNs designed specifically for graph‑level objectives, (2) a narrow focus on a limited subset of architectures (mostly node‑centric models), (3) insufficient diversity of datasets (most benchmarks are chemistry‑centric), (4) an over‑reliance on clean, balanced, and abundant labeled data, and (5) inconsistent experimental pipelines that make reproducibility and fair comparison difficult.

To address these gaps, the authors propose a five‑category taxonomy for graph‑level GNNs:

1. Node‑based GNNs – classic message‑passing networks (GCN, GraphSAGE, GIN, etc.) that aggregate node embeddings with a permutation‑invariant read‑out (mean, sum, max).
2. Hierarchical‑Pooling (HP)‑based GNNs – models that progressively coarsen the graph via learned or similarity‑based cluster assignment matrices (e.g., DiffPool, MinCutPool, TopKPool, SAGPool).
3. Subgraph‑based GNNs – approaches that decompose a graph into overlapping or disjoint subgraphs, learn representations for each subgraph, and then combine them (e.g., ESAN, DropGNN, rooted‑subgraph methods).
4. Graph‑Learning (GL)‑based GNNs – methods that augment the primary task with auxiliary graph reconstruction or transformation objectives, thereby improving representation robustness.
5. Self‑Supervised Learning (SSL)‑based GNNs – pre‑training strategies that exploit unlabeled graphs through contrastive learning, mask‑prediction, or graph‑level consistency losses.

Having defined the taxonomy, the authors introduce OpenGLT, a unified, open‑source evaluation framework that standardizes data splits, hyper‑parameter search, metrics, and reporting across a broad spectrum of scenarios. OpenGLT incorporates 26 datasets spanning four domains (biology, chemistry, social networks, motif graphs), covering 13 classification and 13 regression tasks. Moreover, it automatically generates realistic challenge settings: noisy graphs (random edge/feature perturbations), class‑imbalanced splits, and few‑shot regimes where only a tiny fraction of graphs are labeled.

The experimental campaign evaluates 16 state‑of‑the‑art models (four per category) under identical conditions. Key findings include:

• Node‑based models remain strong baselines on traditional benchmarks but struggle with large, highly structured graphs where hierarchical information is crucial.
• Hierarchical‑pooling models achieve superior performance on large graphs while drastically reducing memory consumption; however, the clustering step can become a computational bottleneck.
• Subgraph‑based models excel when local structural motifs dominate the target (e.g., subgraph counting), but their performance is highly sensitive to the number and size of sampled subgraphs, leading to increased GPU memory demand.
• Graph‑learning models improve robustness to noisy inputs by jointly optimizing reconstruction losses, yielding 2–4 % gains in noisy settings.
• Self‑supervised models provide the most pronounced benefit in low‑label regimes; pre‑training followed by fine‑tuning improves regression accuracy by 3–5 % over purely supervised baselines, and contrastive variants converge faster during fine‑tuning.

Efficiency analysis shows that pooling‑based approaches are the most memory‑efficient, while subgraph‑based methods are the most compute‑intensive due to repeated subgraph processing. SSL models incur a longer pre‑training phase but pay off with rapid downstream adaptation.

The paper also highlights research gaps: subgraph generation currently relies on random deletions or root‑centric sampling, leaving room for domain‑aware, semantics‑driven subgraph extraction; SSL for graph‑level tasks is still nascent, with many possible augmentation strategies unexplored; and scalable clustering algorithms for hierarchical pooling remain an open challenge.

By releasing OpenGLT as an open‑source library and providing detailed documentation, the authors enable the community to benchmark new architectures, extend evaluations to novel domains (e.g., financial transaction networks, geographic graphs), and ensure reproducibility. Overall, the work delivers a comprehensive taxonomy, a rigorously standardized benchmark suite, and actionable insights that together chart a clear roadmap for future advances in graph‑level GNN research.