Research map / Graph learning / GNN architectures
N2GON: Neural Networks for Graph-of-Net with Position Awareness
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.
Proposes Graph-of-Net (GON), where each node is itself a graph, and N2GON, a dual-encoder position-aware network with a PageRank-based constraint network that is evaluated on node classification and biomedical interaction prediction.
Problem
Standard graph learning treats graphs as flat units and misses multi-level structure where each node is itself a graph connected in a larger network, such as proteins in protein-protein interaction networks.
Approach
Dual encoders (one for intra-node-graph structure, one for inter-graph network) plus a label-based constraint network refined by Personalized PageRank scores, used in a contrastive-style loss that pulls together similar node-graph representations according to label agreement and network proximity.
Key results
- On Texas, N2GON reaches 90.16% accuracy versus 87.84% for the best baseline (ACM-GCN), per Table 1.
- On KIBA (DTI), N2GON reports 95.27% ROC-AUC, reported as a 2.66% gain over the top encoder-based baseline.
- Ablation on Cora: removing Grapher II drops accuracy from 89.14% to 87.22%, and removing Grapher I drops it to 85.81%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use N2GON instead of the strongest listed baseline | node classification · Texas, Wisconsin, Actor, Chameleon, Cornell, CiteSeer, Pubmed, and Cora | accuracy | Helped |
| Use N2GON instead of HopGNN | node classification · Squirrel | accuracy: 64.23±1.33 -> 60.98±2.12 | Hurt |
| Use N2GON instead of the strongest listed baseline | biomedical interaction prediction · PPI:HuRI | ROC-AUC: 88.81 ± 4.14 -> 89.83 ± 3.52 | Helped |
| Use N2GON instead of ConjointTriad | biomedical interaction prediction · PPI:HuRI | PR-AUC: 91.01 ± 7.18 -> 91.37 ± 5.23 | Helped |
| Use N2GON instead of ESPF | biomedical interaction prediction · PEPMHC:MHC1 | ROC-AUC: 93.31 ± 5.12 -> 93.77 ± 6.23 | Helped |
| Use N2GON instead of ESPF | biomedical interaction prediction · PEPMHC:MHC1 | PR-AUC: 87.46 ± 6.18 -> 86.94 ± 4.65 | Hurt |
| Use N2GON instead of the strongest listed baseline | biomedical interaction prediction · TCR:Weber | ROC-AUC: 86.51 ± 6.91 -> 89.78 ± 7.38 | Helped |
| Use N2GON instead of Transformer | biomedical interaction prediction · TCR:Weber | PR-AUC: 84.35 ± 3.80 -> 89.75 ± 2.35 | Helped |
| Use N2GON instead of the strongest listed baseline | biomedical interaction prediction · MTI:miRTarBase | ROC-AUC: 89.53 ± 6.71 -> 91.55 ± 7.49 | Helped |
| Use N2GON instead of ConjointTriad | biomedical interaction prediction · MTI:miRTarBase | PR-AUC: 82.82 ± 4.56 -> 84.66 ± 3.95 | Helped |
| Remove Grapher II (ablation) | N2GON · Squirrel | accuracy: 60.98±2.12 -> 59.10±2.48 | Hurt |
| Remove Grapher II (ablation) | N2GON · Cora | accuracy: 89.14±2.01 -> 87.22±1.83 | Hurt |
Surprising findings
- Removing Grapher I (label-identity constraint) hurts more than removing Grapher II (PPR-refined constraint), indicating the label-based component is the main contributor.
Limitations
- Noise in node attributes or connectivity can propagate through encoding layers and degrade performance.
- Results depend on graph construction choices; the paper does not compare multiple construction strategies.
- Benchmark datasets were adapted into GON form by k-hop subgraph sampling per node, which is an assumption rather than a native GON structure.
Experimental setup
Datasets: Cora CiteSeer PubMed Texas Wisconsin Cornell Actor Squirrel Chameleon DAVIS KIBA TWOSIDES
Benchmarks: Cora CiteSeer PubMed Texas Wisconsin Cornell Actor Squirrel Chameleon DAVIS KIBA TWOSIDES
Metrics: accuracy ROC-AUC PR-AUC
Baselines: MLP GCN GAT GraphSAGE GCNII H2GCN ACM-GCN GGCN HopGNN GoGNN AAC ConjointTriad
Builds on: GCN Personalized PageRank HopGNN Geom-GCN
Closest papers by meaning
- Positional Encoding meets Persistent Homology on Graphs (ICML 2025 · similarity 0.86)
- Subgraphormer: Unifying Subgraph GNNs and Graph Transformers via Graph Products (ICML 2024 · similarity 0.86)
- Translating Subgraphs to Nodes Makes Simple GNNs Strong and Efficient for Subgraph Representation Learning (ICML 2024 · similarity 0.86)
- Subgraph Invariant Learning Towards Large-Scale Graph Node Classification (AAAI 2025 · similarity 0.86)
- Learning Efficient Positional Encodings with Graph Neural Networks (ICLR 2025 · similarity 0.86)
- GPEN: Global Position Encoding Network for Enhanced Subgraph Representation Learning (ICML 2025 · similarity 0.86)
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.