Research map / Graph learning / Knowledge and temporal graphs
Heterogeneous Graph Neural Network on Semantic Tree
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 HetTree, a scalable heterogeneous GNN that builds a semantic tree over metapaths, encodes it with subtree attention, and matches pre-computed features with same-metapath labels.
Problem
Existing heterogeneous GNNs ignore the tree hierarchy (prefix parent-child relations) among metapaths and handle features and labels separately or in the same latent space, limiting accuracy and scalability on large heterogeneous graphs.
Approach
Organizes all metapaths up to hop k into a semantic tree, performs offline mean aggregation of features and labels per metapath, and aggregates bottom-up with a subtree attention mechanism that scores children using both parent and sibling-subtree context, while concatenating each metapath's features with its matched labels.
Key results
- On HGB benchmarks HetTree reports Micro-F1 of 95.64 on DBLP, 70.92 on IMDB, 94.19 on ACM and 66.39 on Freebase, beating SeHGNN on all four.
- On Ogbn-Mag HetTree reaches 55.54% test accuracy vs 54.78% for SeHGNN without multi-stage training.
- On the proprietary email graph HetTree reaches 98.48% test accuracy vs 97.26% for SeHGNN.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use semantic-tree aggregation with subtree attention instead of weighted child-sum (ablation) | HetTree · IMDB | Micro-F1: 69.70$\pm$0.24 -> 70.92$\pm$0.29 | Helped |
| use semantic-tree aggregation with subtree attention instead of weighted child-sum (ablation) | HetTree · ACM | Micro-F1: 93.54$\pm$0.23 -> 94.19$\pm$0.20 | Helped |
| use semantic-tree aggregation with subtree attention instead of weighted child-sum (ablation) | HetTree · Ogbn-Mag | Accuracy: 55.38$\pm$0.21 -> 55.54$\pm$0.17 | Helped |
| use subtree attention instead of parent-node attention (ablation) | HetTree · IMDB | Micro-F1: 69.69$\pm$0.41 -> 70.92$\pm$0.29 | Helped |
| use subtree attention instead of parent-node attention (ablation) | HetTree · ACM | Micro-F1: 93.83$\pm$0.18 -> 94.19$\pm$0.20 | Helped |
| use subtree attention instead of parent-node attention (ablation) | HetTree · Ogbn-Mag | Accuracy: 54.87$\pm$0.26 -> 55.54$\pm$0.17 | Helped |
| add training labels as extra inputs (ablation) | HetTree · IMDB | Micro-F1: 70.11$\pm$0.25 -> 70.92$\pm$0.29 | Helped |
| add training labels as extra inputs (ablation) | HetTree · ACM | Micro-F1: 93.41$\pm$0.23 -> 94.19$\pm$0.20 | Helped |
| add training labels as extra inputs (ablation) | HetTree · Ogbn-Mag | Accuracy: 52.93$\pm$0.12 -> 55.54$\pm$0.17 | Helped |
| use HetTree rather than SeHGNN | HGB benchmark models · DBLP, IMDB, ACM, and Freebase | Macro-F1 and Micro-F1 | Helped |
| use HetTree rather than SeHGNN | Ogbn-Mag models · Ogbn-Mag | Accuracy | Helped |
| use HetTree rather than SeHGNN | Email models · Email | Accuracy | Helped |
Surprising findings
- Removing labels (no-label variant) hurts Ogbn-Mag accuracy by about 2.6 points, showing label utilization matters most there.
Limitations
- Evaluated on a proprietary email dataset that cannot be released, limiting reproducibility of that result.
- Relies on offline aggregation over all metapaths up to a user-defined hop k, which can grow with k.
- Ogbn-Mag results are reported without self-enhanced techniques such as multi-stage training.
Experimental setup
Datasets: DBLP IMDB ACM Freebase Ogbn-Mag Email (commercial, proprietary)
Benchmarks: HGB DBLP HGB IMDB HGB ACM HGB Freebase Ogbn-Mag Email
Metrics: Macro-F1 Micro-F1 accuracy ROC curve epoch time memory usage
Baselines: RGCN HAN HetGNN MAGNN HGT HGB SeHGNN NARS LEGNN GAMLP
Builds on: SeHGNN SIGN UniMP GAMLP TreeLSTM
Closest papers by meaning
- HeterSEED: Semantics–Structure Decoupling for Heterogeneous Graph Learning under Heterophily (NeurIPS 2026 · similarity 0.91)
- An Efficient Subgraph-Inferring Framework for Large-Scale Heterogeneous Graphs (AAAI 2024 · similarity 0.89)
- Bootstrapping Heterogeneous Graph Representation Learning via Large Language Models: A Generalized Approach (AAAI 2025 · similarity 0.89)
- HGOT: Self-supervised Heterogeneous Graph Neural Network with Optimal Transport (ICML 2025 · similarity 0.89)
- Contrastive Auxiliary Learning with Structure Transformation for Heterogeneous Graphs (AAAI 2025 · similarity 0.88)
- Harnessing Language Model for Cross-Heterogeneity Graph Knowledge Transfer (AAAI 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.