Research map / Graph learning / Graph self-supervised learning
HGOT: Self-supervised Heterogeneous Graph Neural Network with Optimal Transport
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 HGOT, a self-supervised heterogeneous GNN that aligns optimal-transport plans between meta-path (branch) views and an attention-aggregated central view, removing the need for graph augmentation and positive/negative sampling.
Problem
Contrastive self-supervised learning on heterogeneous graphs requires carefully designed graph augmentations and positive/negative pair selection, and the required similarity level between pairs is hard to determine.
Approach
Build a central aggregated view from meta-path views via attention and a unified adjacency (logical OR); compute fused Gromov-Wasserstein optimal transport plans between each branch graph and the central graph, and train the encoder (HAN backbone) so its representation-space OT plan matches the graph-space plan, avoiding augmentations and positive/negative samples.
Key results
- On node classification HGOT reportedly gains over 6% average accuracy over state-of-the-art methods, and about 34% over GraphSAGE on ACM.
- Node clustering: HGOT shows roughly 2.5% average improvement on DBLP, IMDB and Yelp.
- Ablation: removing the optimal transport plan alignment (replacing with a distance-difference loss) gives the worst results among variants.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use HGOT instead of state-of-the-art baselines | HGOT · Node classification, four datasets | accuracy: more than 6% | Helped |
| Use HGOT instead of baselines | HGOT · Node clustering, DBLP, IMDB, and Yelp | clustering performance: approximately 2.5% | Helped |
| Add aggregated graph generation (ablation) | HGOT · Node classification and clustering, four datasets | performance | Helped |
| Add implicit structure loss (ablation) | HGOT · Node classification and clustering, four datasets | performance | Helped |
| Align optimal transport plans rather than directly optimizing transport distance (ablation) | HGOT · Node classification and clustering, four datasets | performance | Helped |
| Give relatively little contribution to aggregating all semantic information (ablation) | HGOT · Node classification and clustering, four datasets | performance | Helped |
| Increase the implicit structure loss weight ρ up to 1 (ablation) | HGOT · DBLP and IMDB | prediction performance: ρ=1 | Helped |
| Use node attributes only to calculate the transport plan (σ=1) (ablation) | HGOT · DBLP and IMDB | performance: σ=1 | Helped |
| Use both node attributes and edges to calculate the transport plan (ablation) | HGOT · DBLP and IMDB | performance | Hurt |
| Use edge structure only to calculate the transport plan (σ=0) (ablation) | HGOT · DBLP and IMDB | performance: σ=0 | Hurt |
| Vary hidden dimension (ablation) | HGOT · DBLP and IMDB | performance: 64, 128, 256, 512, 1024 | Mixed |
Surprising findings
- Using only node-attribute OT (sigma=1) performs best, while including edge structure in the plan reduces performance, suggesting edges can be partially dropped.
- Removing the aggregated-view module (HGOT-w/o#1) remains competitive, indicating the aggregation contributes little overall.
Limitations
- Fused Gromov-Wasserstein optimization via conditional gradient scales cubically with node count, adding computational cost.
- Performance is sensitive to trade-off hyperparameters rho and sigma; edge-based OT alone performs poorly.
- Evaluated only on four datasets with few meta-paths.
Experimental setup
Datasets: DBLP ACM IMDB Yelp
Benchmarks: DBLP ACM IMDB Yelp
Metrics: Micro-F1 Macro-F1 ACC NMI ARI
Baselines: GCN GAT HAN GraphSAGE GAE DGI DeepWalk Metapath2vec HERec HDGI HEAD HeCo
Builds on: HAN GALOPA HeCo Gromov-Wasserstein distance
Closest papers by meaning
- HGPrompt: Bridging Homogeneous and Heterogeneous Graphs for Few-Shot Prompt Learning (AAAI 2024 · similarity 0.89)
- HeterSEED: Semantics–Structure Decoupling for Heterogeneous Graph Learning under Heterophily (NeurIPS 2026 · similarity 0.89)
- Bootstrapping Heterogeneous Graph Representation Learning via Large Language Models: A Generalized Approach (AAAI 2025 · similarity 0.89)
- Heterogeneous Graph Neural Network on Semantic Tree (AAAI 2025 · similarity 0.89)
- Contrastive Auxiliary Learning with Structure Transformation for Heterogeneous Graphs (AAAI 2025 · similarity 0.89)
- HG-Adapter: Improving Pre-Trained Heterogeneous Graph Neural Networks with Dual Adapters (ICLR 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.