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

1. Introduction

• Prior work had not established comparable classic-GNN performance for graph-level tasks, and an earlier study was limited to LRGB and still trailed GTs overall.The authors frame this gap as the question of whether classic GNNs can excel on graph-level tasks.Source: §1 · p. 2
• GNN+ integrates edge features, normalization, dropout, residual connections, feed-forward networks, and positional encoding into message passing, with each technique treated as a tunable hyperparameter.The study evaluates GCN, GIN, and GatedGCN enhanced by this framework.Source: §1 · p. 2
• Across 14 graph-level datasets, enhanced classic GNNs match or outperform state-of-the-art GTs, rank in the top three on every dataset, and place first on eight.They also run several times faster than GTs on many datasets.Source: §1 · p. 2
• Edge features are particularly beneficial on molecular and image-superpixel datasets, where they encode domain-specific information.Source: §1 · p. 2
• Dropout benefits most graph-level datasets, and a modest dropout rate is reported as sufficient and optimal.Source: §1 · p. 2
• Residual connections are generally essential, except for shallow GNNs applied to small graphs.Source: §1 · p. 2
• Feed-forward networks are especially pivotal for simpler models such as GCN in graph-level tasks.Source: §1 · p. 2
• Positional encoding is more critical for small-scale datasets than for large-scale ones.The authors describe the roles of the integrated techniques as nuanced and relevant to future GNN design and application.Source: §1 · p. 2

2. Classic GNNs for Graph-level Tasks

• Message-passing GNNs update each node representation by aggregating information from neighboring nodes, then produce a final node representation after the last layer.The section focuses on GCN, GIN, and GatedGCN, which differ in how they learn node representations.Source: §2 · pp. 2–3
• GCN is presented as a classic graph convolution model, with its layer update using a trainable weight matrix and activation function.Source: §2 · p. 3
• GIN learns node representations using an update with a multi-layer perceptron, usually consisting of two layers.Source: §2 · p. 3
• GatedGCN incorporates gating into graph convolution; its gate determines how much each neighboring node contributes to the current node's update.Source: §2 · p. 3
• Graph-level datasets associate each graph with classification labels or regression targets, and prediction uses a readout over node representations followed by a prediction head.For graph prediction, the readout often uses global mean pooling; training minimizes loss over graphs in the training set.Source: §2 · p. 3

3.1. Edge Feature Integration

• Edge features are integrated into message passing through a learned transformation of each edge feature alongside the source-node representation.For GCN, the edge feature vector is multiplied by a trainable layer-specific weight matrix.Source: §3.1 · pp. 3–4

3.2. Normalization

• Batch Normalization is applied to layer outputs before the activation function.The section describes normalization as stabilizing training by mitigating changes in node-embedding distributions across layers.Source: §3.2 · p. 4

3.3. Dropout

• Dropout is applied to node embeddings after activation.The section relates dropout in GNNs to reducing co-adaptation effects propagated through message passing.Source: §3.3 · p. 4

3.4. Residual Connection

• Residual connections allow deeper GNNs, from 3 to 20 layers, to achieve strong performance.The section contrasts this range with typical GNN depths of 2 to 5 layers and notes the challenge of over-smoothing in deeper models.Source: §3.4 · p. 4

3.5. Feed-Forward Network

• GNN+ appends a fully connected feed-forward network at the end of each GNN layer.The FFN is described as adding complex feature transformations and nonlinearity.Source: §3.5 · p. 4

3.6. Positional Encoding

• GNN+ uses Random Walk Structural Encoding as its positional encoding method to improve GNN performance.The section describes positional encodings as incorporating graph positional or structural information, typically by adding or concatenating them to input node features.Source: §3.6 · pp. 4–5

Datasets

• The evaluation covers 14 graph-level datasets from GNN Benchmark, LRGB, and OGB, using their standard splits and metrics.The datasets include graph benchmarks, long-range graph tasks, and OGB tasks.Source: §4 · p. 5

Baselines

• The baselines include classic GNNs, state-of-the-art Graph Transformers, and graph state-space models.Reported baseline results primarily come from prior benchmark studies, with others taken from original papers or official leaderboards when possible.Source: §4 · pp. 5–6

Datasets

• Table 3 reports test performance on five LRGB datasets and lists approximately 500K parameters for all models.Its caption says baseline results come from the respective original papers.Source: Table 3 · p. 6

Hyperparameter Configurations

• Hyperparameter tuning covers learning rates, 3–20 layers, and GNN+ modules, but is not exhaustive.The authors also retrained baseline GTs using the same search space and training environments, then reported original-paper results because retraining did not surpass them; GNN+ results are means and standard deviations over five runs.Source: §4 · p. 6

5.1. Overall Performance

• Across 14 datasets, enhanced classic GNNs ranked top three throughout, took first place in eight, and showed superior efficiency.Source: §5.1 · p. 6
• GNN+ improves ZINC, PATTERN, and CLUSTER; GatedGCN+ beats SOTA models on MNIST and CIFAR and ranks third on PascalVOC-SP and COCO-SP, while GCN+ excels on peptide datasets.Source: §5.1 · pp. 6–7
• On ogbg-ppa, GatedGCN+ improves performance by approximately 9% and ranks first; it ranks top three across OGB and tops three of four datasets.Source: Table 4 · p. 7, §5.1 · p. 7

5.2. Ablation Study

• Ablations broadly penalize component removal; edge features matter for molecular and image-superpixel structure, while normalization matters more on larger datasets, dropping ogbg-ppa accuracy around 15%.Source: §5.2 · pp. 7–8
• Dropout helps most GNN Benchmark, LRGB, and large OGB datasets: 97% of optimal rates are ≤0.2 and 64% are ≤0.1.Source: §5.2 · p. 8
• Removing residuals usually harms results, but may not help shallow 3–5-layer peptide models; deeper networks on larger graphs rely on them for gradient flow and stable long-range exchange.Source: §5.2 · pp. 8–9
• Removing FFN causes broad GCN+ and GIN+ declines, including an 88% relative accuracy drop for GIN+ on MNIST; GatedGCN+ relies less on FFN, and peptide models omit it.Source: §5.2 · p. 9
• Positional encoding substantially benefits several smaller datasets with 10,000–40,000 graphs, while its impact is less pronounced on datasets exceeding 100,000 graphs.Source: §5.2 · p. 9

• Node-level benchmark audits found that claimed architecture gains can disappear under fair evaluation; Luo et al. (2024b) reported classic GNNs matching or surpassing SOTA across 18 node-classification datasets.Source: §6 · p. 9
• Graph-level evaluation work includes OGB standardization, fixed-budget comparisons across diverse graphs, and LRGB benchmarks designed to assess long-range dependencies.Source: §6 · p. 9
• Tönshoff et al. (2023) found classic GNNs can close the gap with GTs under tuning, while prior studies were limited by dataset scope, diversity, or hyperparameter coverage.Source: §6 · p. 9

• The study offers empirical evidence on existing benchmarks but no theoretical explanation for the enhanced GNNs’ observed advantage over GTs.Source: §7 · p. 9
• Future work should explain how GNN+ components and their interactions affect expressiveness, representation, oversmoothing, and oversquashing.Source: §7 · p. 9

• Across 14 benchmark datasets, GNN+ enhanced GCN, GIN, and GatedGCN to match or outperform SOTA GTs with greater computational efficiency.Source: §8 · p. 9

Impact Statement

• The authors describe the work as an empirical benchmarking study that provides practical guidance for future model selection and design, with no immediate societal consequences to highlight.The impact statement says the findings challenge common assumptions about classic GNN performance on graph-level tasks.Source: Impact Statement · p. 10

A.1. Computing Environment

• Experiments were conducted on a single workstation equipped with 8 RTX 3090 GPUs.Source: §A.1 · p. 14

A.2. Datasets

• GNN Benchmark includes ZINC molecular solubility regression and 10-class graph classification on MNIST and CIFAR10 images represented as superpixel graphs.The suite also includes synthetic PATTERN and CLUSTER datasets for subgraph-pattern recognition and cluster-ID inference.Source: §A.2 · p. 14
• LRGB covers peptide functional-class classification and structural-property regression, superpixel object-class node classification, and MalNet-Tiny software-type prediction from function-call graphs.PCQM-Contact was excluded because its download link was no longer valid.Source: §A.2 · p. 14
• OGB adds larger-scale molecular-property, protein-network origin, and code-function-name prediction datasets, with graphs ranging from hundreds of thousands to millions.The tasks include HIV-inhibition classification, 128-assay prediction, taxonomic-group classification, and prediction of the first five function-name subtokens.Source: §A.2 · p. 14

A.3. Hyperparameters and Reproducibility

• Hyperparameter settings are detailed in Tables 8–13, the validation set selects the best settings, and the code is released under the MIT License.The study follows GraphGPS experiment settings and points to its code for further hyperparameter details.Source: §A.3 · pp. 14–15

B.1. Hyperparameter Sensitivity Analysis

• GCN+ and GatedGCN+ peak at 12 layers on PATTERN, while CLUSTER and PascalVOC-SP improve with depth; without residuals, performance degrades especially beyond 8 layers.The degradation is most pronounced on CLUSTER and PascalVOC-SP, where the authors attribute it to over-smoothing and over-squashing.Source: §B.1 · pp. 17–18, Figure 3 · p. 18
• Dropout rates of 0.1–0.2 are optimal across four datasets, while rates of 0.3 or higher significantly degrade performance.On Peptides-struct, MAE rises sharply beyond 0.2 dropout; on CLUSTER, accuracy collapses at 0.5.Source: §B.1 · p. 18, Figure 4 · p. 19

B.2. Performance of a Fixed GNN+ Configuration

• A fixed GNN+ configuration integrating all six techniques performs comparably to the best dataset-tuned configurations across nine datasets.The authors present this as evidence of practicality and robustness without extensive tuning.Source: §B.2 · p. 18, Table 14 · p. 18

B.3. Additional Baselines

• GCN+ matches or exceeds LDP and HRN accuracy on five small datasets; on two OGB datasets, those linear-time baselines train faster but perform significantly worse.Training times are negligible across models on the small datasets.Source: §B.3 · pp. 18–19, Table 15 · p. 19, Table 16 · p. 19
• GCN+ consistently outperforms MTGCN and CO-GNN on Peptides-func and Peptides-struct.Source: §B.3 · pp. 19–20, Table 17 · p. 20

B.4. Visualization

• On Peptides-func, GCN+ embeddings show more distinct class separation and larger inter-class margins than vanilla GCN embeddings.The comparison is a t-SNE visualization.Source: §B.4 · p. 20, Figure 5 · p. 20

B.5. Static Attention Problem of GTs

• The authors identify a potential static-attention limitation in representative GTs: attention appears query-invariant, and a misclassified ZINC molecule illustrates low attention to functionally significant substructures.In that GraphGPS example, nodes focus mainly on pyrazoline and benzene rings rather than the C–O–N=N and nitro groups; the authors connect this pattern to insufficient subgraph sensitivity and adverse prediction accuracy.Source: §B.5 · p. 20, Figure 6 · p. 21

C.1. Clarification on GNN+ and Classic GNNs

• GNN+ unifies widely used components while preserving message passing; the authors describe it as a practical extension, not a departure from the message-passing paradigm.The components are not novel or specific to one architecture.Source: §C.1 · p. 21