KE-HGNN

# Knowledge-Embedded Hypergraph Neural Networks **KE-HGNN** · Official Implementation [![TPAMI](https://img.shields.io/badge/IEEE-TPAMI%202026-9b59b6?style=flat-square)](https://doi.org/10.1109/TPAMI.2026.3674800) [![Python 3.11](https://img.shields.io/badge/Python-3.11-3776AB?style=flat-square&logo=python&logoColor=white)](https://www.python.org/) [![PyTorch](https://img.shields.io/badge/PyTorch-%E2%89%A52.0-EE4C2C?style=flat-square&logo=pytorch&logoColor=white)](https://pytorch.org/) [![DHG](https://img.shields.io/badge/DHG-0.9.7-2ea44f?style=flat-square)](https://github.com/iMoonLab/DeepHypergraph) [![License: MIT](https://img.shields.io/badge/License-MIT-f1c40f?style=flat-square)](LICENSE) *Yifan Feng, Yifan Zhang, Shaoyi Du, Shihui Ying\*, Zongze Wu, Yue Gao* *\* Corresponding author* KE-HGNN Framework

**The KE-Phase embeds high-order structural knowledge (HOI-Encoder) and task-driven rule knowledge (TDR-Encoder) into vertex features, fused by MDK-Fusion before the HGNN backbone.**

🎯 Overview

Conventional HGNNs rely solely on structure-guided feature smoothing. KE-HGNN instead explicitly extracts and embeds two kinds of knowledge before message passing:

HOI-Encoder (High-Order Incidence Encoder, Eq. 3) — a permutation-invariant map of the incidence matrix producing structure-related embeddings: Xs[v] = σ(w · Σe H[v,e] + b).
TDR-Encoder (Task-Driven Rule Encoder, Eq. 4–8) — parses a pre-trained GBDT forest into rule embeddings, with depth-aware weighting ψp(o) = α / (d(o) + β) (Eq. 6).
MDK-Fusion (Eq. 9) — multi-dimensional knowledge fusion: Z = Xr ‖ (Xs ⊙ Xr).

Two instantiations:

`nn_method`	Model	Definition
`RD_HGNN`	RD-HGNN	`HGNN(r(X, T), H)` — rule knowledge only (Eq. 10)
`DD_HGNN`	DD-HGNN	rule + structure dual knowledge with MDK-Fusion

📦 Installation

Python 3.11 is recommended. All datasets are downloaded automatically by the
DHG library on first use.

pip install -r requirements.txt

For GPU support, install a CUDA build of PyTorch first, e.g.:

pip install torch --index-url https://download.pytorch.org/whl/cu121

🚀 Quick Start

# single run (default: Cora, RD-HGNN)
python train.py

# paper setting: 5 seeds, mean ± std
python run_multi_seed.py

# switch to DD-HGNN
python run_multi_seed.py model.nn_method=DD_HGNN

All hyperparameters live in configs/trans_config.yaml and can be overridden
from the command line via hydra.

📊 Results

Transductive vertex classification (accuracy %, 5 seeds, Table V in the paper):

Dataset	MLP	GCN	GBDT	HGNN	RD-HGNN	DD-HGNN
Cora	69.04 ± 2.04	83.63 ± 1.12	74.30 ± 1.67	81.81 ± 2.51	88.64 ± 1.34	87.22 ± 1.30
CA-Cora	68.32 ± 2.90	75.23 ± 2.05	71.54 ± 2.08	75.67 ± 2.38	79.36 ± 2.60	79.75 ± 2.10
CC-Cora	68.55 ± 3.30	73.21 ± 1.22	71.69 ± 2.12	73.41 ± 2.16	74.30 ± 4.09	76.06 ± 3.48
DBLP-Conf	79.24 ± 0.47	90.57 ± 0.81	79.77 ± 1.62	93.79 ± 0.40	94.68 ± 0.22	94.84 ± 0.50

Reproduce

bash reproduce/run_core_datasets.sh   # all four datasets × {RD-HGNN, DD-HGNN}

or individually (hyperparameters from Table IV of the paper):

# Cora (graph)
python run_multi_seed.py data.name=cora model.n_est=60  model.max_depth=4 model.alpha=10 model.beta=6  model.xgb_lr=0.6 model.nn_method=RD_HGNN
# CA-Cora (coauthorship hypergraph)
python run_multi_seed.py data.name=coauthorship_cora model.n_est=40  model.max_depth=5 model.alpha=7 model.beta=8  model.xgb_lr=0.4 model.nn_method=RD_HGNN
# CC-Cora (cocitation hypergraph)
python run_multi_seed.py data.name=cocitation_cora model.n_est=140 model.max_depth=3 model.alpha=9 model.beta=8  model.xgb_lr=0.4 model.nn_method=RD_HGNN
# DBLP-Conf (hypergraph)
python run_multi_seed.py data.name=dblp4k-conf model.n_est=100 model.max_depth=3 model.alpha=9 model.beta=10 model.xgb_lr=0.4 model.nn_method=RD_HGNN

The transductive split is 60% train / 25% val / 15% test (stratified), matching
Table II of the paper; graph datasets use their original edges directly as
size-2 hyperedges (cf. Table I).

<details>
<summary><b>Available datasets</b> (via DHG)</summary>

Graphs: cora, pubmed, citeseer.
Hypergraphs: coauthorship_cora, coauthorship_dblp, cocitation_cora,
cocitation_pubmed, cocitation_citeseer, dblp4k-paper, dblp4k-term,
dblp4k-conf, imdb4k, news20, recipe-100k, recipe-200k.

</details>

🗂 Project Structure

├── configs/trans_config.yaml   # hydra config (per-dataset hyperparameters in comments)
├── ke_hgnn/
│   ├── model.py                # KE_HGNN estimator + NeuralNet (RD_HGNN / DD_HGNN backbones)
│   ├── tdr_encoder.py          # TDR-Encoder
│   └── utils.py                # data loading, splits, hypergraph helpers, multi-seed metric
├── train.py                    # single-seed entry
├── run_multi_seed.py           # 5-seed mean ± std entry
└── reproduce/run_core_datasets.sh

📝 Citation

@article{feng2026kehgnn,
  title   = {Knowledge-Embedded Hypergraph Neural Networks},
  author  = {Feng, Yifan and Zhang, Yifan and Du, Shaoyi and Ying, Shihui and Wu, Zongze and Gao, Yue},
  journal = {IEEE Transactions on Pattern Analysis and Machine Intelligence},
  volume  = {48},
  number  = {7},
  pages   = {8613--8625},
  month   = jul,
  year    = {2026},
  doi     = {10.1109/TPAMI.2026.3674800}
}

📄 License

This project is released under the MIT License.

Specs

README