Research map / Graph learning / Graph self-supervised learning
iN2V: Bringing Transductive Node Embeddings to Inductive Graphs
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 iN2V, a post-hoc iterative averaging procedure that extends node2vec embeddings to unseen test nodes in the inductive setting, combined with loss- and sampling-based modifications to N2V training.
Problem
Shallow transductive node embeddings like node2vec cannot produce embeddings for new nodes unseen during training, which limits them to transductive settings and leaves nodes with missing features without structural information.
Approach
Extend trained N2V embeddings to unseen nodes by iteratively setting each new node to the mean of its already-embedded neighbors while letting training embeddings adapt via a convex combination; N2V training is modified with a neighborhood-closeness loss, a neighbor-diversity loss, and a sampling scheme that replaces embeddings with neighborhood means so embeddings suit this post-hoc extension.
Key results
- iN2V improves node classification by about 1 point on average over Feature Propagation and up to 6 points depending on dataset and number of unseen nodes (abstract).
- With GraphSAGE on Cora at 10% training data, iN2V reaches 78.88% vs 77.91% for Feature Propagation and 42.18% for plain inductive N2V.
- On Computers at 10% training, iN2V reaches 88.36% vs 82.79% for Feature Propagation and 77.64% for inductive N2V; iN2V embeddings beat original features on some low-data splits.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use adaptive iN2V post-hoc embedding extension instead of Feature Propagation | N2V embeddings with MLP · All datasets and splits | accuracy: 62.97 -> 64.38 | Helped |
| Use adaptive iN2V post-hoc embedding extension instead of Feature Propagation | N2V embeddings with GraphSAGE · All datasets and splits | accuracy: 63.20 -> 63.75 | Helped |
| Allow training embeddings to adapt during post-hoc extension rather than freezing them (ablation) | iN2V with MLP · All datasets and splits | accuracy: 61.74 -> 64.38 | Helped |
| Allow training embeddings to adapt during post-hoc extension rather than freezing them (ablation) | iN2V with GraphSAGE · All datasets and splits | accuracy: 62.88 -> 63.75 | Helped |
| Add loss-based N2V training modification (ablation) | iN2V with MLP · All datasets and splits | accuracy: 64.38 -> 64.57 | Helped |
| Add loss-based N2V training modification (ablation) | iN2V with GraphSAGE · All datasets and splits | accuracy: 63.75 -> 63.95 | Helped |
| Add sampling-based N2V training modification (ablation) | iN2V with MLP · All datasets and splits | accuracy: 64.38 -> 64.42 | No clear effect |
| Add sampling-based N2V training modification (ablation) | iN2V with GraphSAGE · All datasets and splits | accuracy: 63.75 -> 63.83 | Helped |
| Add N2V embeddings alongside original graph features instead of using features alone | MLP · All datasets and splits | accuracy: 71.01 -> 73.99 | Helped |
| Add N2V embeddings alongside original graph features instead of using features alone | GraphSAGE · All datasets and splits | accuracy: 74.73 -> 74.24 | Hurt |
Surprising findings
- On heterophilic Actor and Roman-empire datasets, N2V-based models essentially predict the majority class, so N2V embeddings add no useful signal there.
- MLP outperforms GraphSAGE when using post-hoc iN2V embeddings, since the post-hoc extension already performs neighborhood aggregation.
Limitations
- Only nodes with a path to at least one training node receive embeddings; disconnected components without training nodes are unsupported.
- Performance on heterophilic datasets remains near random because N2V captures homophilic structure.
- Random splits leave some training nodes unconnected at small training fractions.
Experimental setup
Datasets: Cora CiteSeer PubMed Computers Photo WikiCS Actor Amazon-ratings Roman-empire
Benchmarks: Cora CiteSeer PubMed Computers Photo WikiCS Actor Amazon-ratings Roman-empire
Metrics: accuracy
Baselines: N2V (inductive) Feature Propagation Original features N2V (transductive) GAT GIN
Builds on: node2vec Feature Propagation GraphSAGE GLNN
Closest papers by meaning
- Learning Representations for Hierarchies with Minimal Support (NeurIPS 2024 · similarity 0.88)
- Node Identifiers: Compact, Discrete Representations for Efficient Graph Learning (ICLR 2025 · similarity 0.87)
- Subgraph Invariant Learning Towards Large-Scale Graph Node Classification (AAAI 2025 · similarity 0.85)
- Weighted Embeddings for Low-Dimensional Graph Representation (AAAI 2025 · similarity 0.85)
- Community Detection Guarantees using Embeddings Learned by Node2Vec (NeurIPS 2024 · similarity 0.85)
- Translating Subgraphs to Nodes Makes Simple GNNs Strong and Efficient for Subgraph Representation Learning (ICML 2024 · similarity 0.85)
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.