Research map / Graph learning / GNN architectures
The Intelligible and Effective Graph Neural Additive Network
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces GNAN, a graph neural additive network that extends generalized additive models to graphs by learning distance and feature shape functions, giving interpretable-by-design global and local explanations with accuracy comparable to black-box GNNs.
Problem
GNNs are black-box models requiring post-hoc explanations, which are insufficient in high-stakes settings where transparency is needed; an intelligible graph model with competitive accuracy is lacking.
Approach
Extend GAMs to graphs: each node's k-th feature passes through a shape function f_k, and nodes are weighted by a learned distance function rho of 1/(1+shortest-path distance), normalized by the number of nodes at that distance; the model is fully described by a few univariate plots and avoids message passing.
Key results
- GNAN was best or second-best in 9 of 13 tasks compared against GraphConv, GraphSAGE, GIN, GATv2, GTransformer and FSGNN.
- On long-range QM9 tasks, GNAN achieved the lowest MAE on mu (2.55) and alpha_HOMO (1.40) among evaluated models, and on alpha (4.28) it was second only to GATv2 4.39 and GTransformer 4.30 within the table's ranking.
- GNAN reached 85.7% on Cornell and 86.9% on PubMed, matching or exceeding black-box GNNs on several node tasks.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GNAN instead of FSGNN | node classification · Cornell | accuracy: 86.0 ± 4.1 -> 85.7 ± 4.8 | Hurt |
| Use GNAN instead of GATv2 | node classification · Tolokers | ROC-AUC: 83.8 ± 1.1 -> 84.5 ± 0.9 | Helped |
| Use GNAN instead of GATv2 | node classification · Cora | accuracy: 83.1 ± 0.9 -> 81.1 ± 1.5 | Hurt |
| Use GNAN instead of GIN | node classification · Citeseer | accuracy: 77.1 ± 1.9 -> 75.8 ± 0.6 | Hurt |
| Use GNAN instead of GraphSAGE | node classification · PubMed | accuracy: 88.4 ± 0.4 -> 86.9 ± 1.2 | Hurt |
| Use GNAN instead of GATv2 | node classification · ogb-arxiv | accuracy: 74.0 ± 2.1 -> 74.1 ± 1.5 | Helped |
| Use GNAN instead of GIN | molecular property regression · µ | MAE: 2.60 ± 0.1 -> 2.55 ± 0.1 | Helped |
| Use GNAN instead of GTransformer | molecular property regression · α | MAE: 4.30 ± 0.5 -> 4.28 ± 0.9 | Helped |
| Use GNAN instead of GATv2 | molecular property regression · αHOMO | MAE: 1.41 ± 0.1 -> 1.40 ± 0.1 | Helped |
| Use GNAN instead of GTransformer | graph classification · Proteins | accuracy: 73.9 ± 1.5 -> 73.2 ± 3.1 | Hurt |
| Use GNAN instead of GTransformer | graph classification · Mutagen | accuracy: 73.1 ± 0.9 -> 72.2 ± 1.0 | Hurt |
| Use GNAN instead of GraphSAGE | graph classification · PTC | accuracy: 67.1±12.6 -> 64.9 ± 7.1 | Hurt |
Surprising findings
- A limited-capacity additive model matches more expressive GNNs on several real-world graph datasets, suggesting some graph problems are simple.
Limitations
- Reduced capacity may limit accuracy on tasks needing complex feature interactions; the model restricts cross-products of features and graph topology.
- Shape functions with ReLU can be step-like (e.g., PubMed), so smoothness may require spline activations.
- Scaling to many features may require regularization to limit the number of shape functions.
Experimental setup
Datasets: Mutagenicity PubMed Cora Citeseer ogb-arxiv Cornell Tolokers NCI1 Proteins PTC QM9
Benchmarks: Cornell Tolokers Cora Citeseer PubMed ogb-arxiv Proteins Mutagenicity PTC NCI1 QM9 mu QM9 alpha
Metrics: accuracy ROC-AUC MAE
Baselines: GraphConv GraphSAGE GIN GATv2 GTransformer FSGNN
Builds on: Neural Additive Models Generalized Additive Models Kolmogorov-Arnold Networks
Closest papers by meaning
- Towards Robust Fidelity for Evaluating Explainability of Graph Neural Networks (ICLR 2024 · similarity 0.89)
- Self-Interpretable Graph Learning with Sufficient and Necessary Explanations (AAAI 2024 · similarity 0.88)
- Reconsidering Faithfulness in Regular, Self-Explainable and Domain Invariant GNNs (ICLR 2025 · similarity 0.88)
- From GNNs to Trees: Multi-Granular Interpretability for Graph Neural Networks (ICLR 2025 · similarity 0.88)
- Stratified GNN Explanations through Sufficient Expansion (AAAI 2024 · similarity 0.88)
- Can Classic GNNs Be Strong Baselines for Graph-level Tasks? Simple Architectures Meet Excellence (ICML 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.