Research map / Graph learning / Knowledge and temporal graphs
Deep Structural Knowledge Exploitation and Synergy for Estimating Node Importance Value on Heterogeneous Information Networks
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 SKES, which estimates HIN node importance values by embedding metapath-induced structural knowledge (centrality and similarity) and measuring 1-Wasserstein distance of each node's feature distribution to a random reference.
Problem
Estimating node importance values (not just rankings) on heterogeneous information networks, where isolated node value prediction by fully adaptive GNNs underuses structural knowledge and offers limited interpretability.
Approach
Treat node importance as the 1-Wasserstein distance between the node's feature distribution and a fixed uninformative random reference distribution, computed via a sorting-based optimal transport map over self-attention representations fused from metapath-induced centrality and similarity knowledge.
Key results
- SKES reduces MAE, RMSE and NRMSE versus the best competing method by 2.75%, 4.76% and 1.20% on average across datasets, with the largest gain of 15.09% on MUSIC10K Song MAE.
- On importance-based ranking, SKES improves SPEARMAN by 3.78% to 19.96% over baselines on most target types, e.g., 0.589 vs 0.491 for DBLP Author.
- Removing the Wasserstein-based estimator (w/o WD) or the learnable lambda vector degrades Movie MSE from 0.509 to 0.540 and 0.524 respectively.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SKES instead of the strongest listed baseline | Node importance estimation · MUSIC10K Song | MAE: 0.122 -> 0.106 | Helped |
| Use SKES instead of the strongest listed baseline | Node importance estimation · TMDB5K Movie | RMSE: 0.664 -> 0.667 | Hurt |
| Use SKES instead of the strongest listed baseline | Node importance estimation · TMDB5K Director | MAE: 0.268 -> 0.274 | Hurt |
| Add marginal ranking loss (ablation) | SKES · Node ranking (dataset not specified) | SPEARMAN: 0.680 -> 0.755 (+11.03%) | Helped |
| Add marginal ranking loss (ablation) | SKES · Node ranking (dataset not specified) | NDCG: 0.847→0.902 (+6.49%) | Helped |
| Remove the Wasserstein distance and use a two-layer MLP (ablation) | SKES · TMDB5K Movie | RMSE: 0.667 -> 0.692 | Hurt |
| Remove the learnable regression vector λ (ablation) | SKES · TMDB5K Movie | RMSE: 0.675 -> 0.667 | Hurt |
| Remove the learnable regression vector λ (ablation) | SKES · TMDB5K Director | NRMSE: 0.092 -> 0.083 | Hurt |
| Remove explicit network heterogeneity learning and treat the network as homogeneous (ablation) | SKES · TMDB5K Movie | MSE: 0.509 -> 0.551 | Hurt |
| Disable self-attention and use the aggregated embedding directly (ablation) | SKES · TMDB5K Movie | RMSE: 0.667 -> 0.696 | Hurt |
| Disable self-attention and use the aggregated embedding directly (ablation) | SKES · TMDB5K Director | RMSE: 0.507 -> 0.549 | Hurt |
| Remove 40% of structural prior knowledge (ablation) | SKES · TMDB5K Movie | MAE | Hurt |
Surprising findings
- On TMDB5K Director, SKES underperforms HIVEN on MAE (-2.19%), which the authors attribute to data scarcity (1.5% of nodes) producing outliers.
- Adding a marginal ranking loss improves SPEARMAN by 11.03% but increases training time per epoch by 63.62%.
Limitations
- Performance gap on the Director type of TMDB5K attributed to scarce training data.
- Ranking-loss variant adds substantial training-time overhead.
Experimental setup
Datasets: MUSIC10K TMDB5K DBLP
Benchmarks: MUSIC10K TMDB5K DBLP
Metrics: MAE RMSE NRMSE NDCG SPEARMAN
Baselines: PageRank personalized PageRank linear regression random forest GAT HGT GENI MULTIIMPORT RGTN HIVEN
Builds on: HIVEN GENI PathSim node2vec HGT
Closest papers by meaning
- Uncertainty Estimation for Heterophilic Graphs Through the Lens of Information Theory (ICML 2025 · similarity 0.87)
- HeterSEED: Semantics–Structure Decoupling for Heterogeneous Graph Learning under Heterophily (NeurIPS 2026 · similarity 0.85)
- Toward a Unified Statistical Theory of Unsupervised Pretraining and Supervised Neural Knowledge Graph Learning (NeurIPS 2026 · similarity 0.84)
- Knowledge Graphs Can be Learned with Just Intersection Features (ICML 2024 · similarity 0.84)
- GRAIN: Multi-Granular and Implicit Information Aggregation Graph Neural Network for Heterophilous Graphs (AAAI 2025 · similarity 0.84)
- On the Impact of Feature Heterophily on Link Prediction with Graph Neural Networks (NeurIPS 2024 · similarity 0.84)
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.