HGNN

This repository aims to reproduce HGCN proposed in the paper entitled "Hypergraph Neural Networks (AAAI'2019)". We refer to the original repository of HGNN to implement this repository.

Dependencies

We use the following Python packages to implement this. Install them in your environment using pip install -r requirements.txt.

• tqdm == 4.63.0
• torch == 1.4.0
• scipy == 1.5.2
• numpy == 1.19.2

Training

To train and evaluate the model, type the following command:

python train.py

It will train the model for the node classification task in the given hypergraph.

In this project, you can control the value of each hyperparameter in config.yaml. The details of the hyperparameters are in the following table.

Option	Description	Default
on_dataset	cora, pubmed, modelnet40, NTU	modelnet40
K_neigs	assignmetn to bulid hyperedge in modelnet40, NTU dataset	10
n_hid	feature numbers of hidden layer	128
dropout	dropout probability for hidden layers	0.5
max_epoch	number of training epochs	600
lr	learning rate	0.001
decay_rate	weight decay	0.0005
decay_step	weight decay step	200
gamma	gamma	0.9

• Visual object task option
  Select the feature that contribute to construct hypregraph incidence matrix by changing the status of parameters.

# config/config.yaml
use_mvcnn_feature_for_structure: True
use_gvcnn_feature_for_structure: True
use_mvcnn_feature: False
use_gvcnn_feature: True

Experimental setting

Datasets

We use the following datasets for the experiments.

• coauthorship : dblp, cora
• cocitation : citeseer, cora, pubmed
• visual object : MVCNN feature, GVCNN featrue

Default Hyperparameters

The followings are default hyperparameters, as described in the original repository.

• hidden nodes size per layer : 128
• dropout rate : 0.5
• learning rate : 0.001
• decay rate : 0.0005
• decay step : 200
• training epochs : 600

How to construct hypergraph from graph dataset CORA, Pumbed

The originaly repository does not support hypergraph dataset CORA, Pumbed. The way to produce the hypergraph dataset for its paper experiment(the related issue is found here).

Becareful NaN or Inf problem during processing(the related issue is found here)

Or download dataset from here.

The experimental result of the paper

Dataset	Pubmed	Cora
HGNN	80.1%	81.6%

• Dataset : ModelNet40

Features for structure	Visual object MVCNN feature+GVCNN featrue	Visual object MVCNN feature	Visual object GVCNN featrue
GVCNN	92.6%	91.8%	96.6%
MVCNN	92.5%	91.0%	96.6%
GVCNN+MVCNN	-	-	96.7%

• Dataset : NTU

Features for structure	Visual object MVCNN feature+GVCNN featrue	Visual object MVCNN feature	Visual object GVCNN featrue
GVCNN	82.5%	79.1%	84.2%
MVCNN	77.2%	75.6%	83.6%
GVCNN+MVCNN	-	-	84.2%

The reproduced result

Dataset	Pubmed	Cora
HGNN	-	-

• Dataset : ModelNet40

Features for structure	Visual object MVCNN feature+GVCNN featrue	Visual object MVCNN feature	Visual object GVCNN featrue
GVCNN	-	-	-
MVCNN	-	-	-
GVCNN+MVCNN	-	-	-

• Dataset : NTU

Features for structure	Visual object MVCNN feature+GVCNN featrue	Visual object MVCNN feature	Visual object GVCNN featrue
GVCNN	-	-	-
MVCNN	-	-	-
GVCNN+MVCNN	-	-	-

Hyperparameter tuning

We think that the problematic hyperparameters are - and -. To find better hyperparameters, we perform grid searches for - and - as follows:

• [-] = [-]
• [-] = [-]

We summarize the result of this experiment in the following table. We report the average of test errors (lower is better) with their standard deviation over 10 runs (accuracy = 100 - error).

Task	Method	Learning rate	Epochs	Error ± std

Recommand a deep learning toolbox named DHG for GNN and HGNN

Follow link here

References

[1] Feng, Y., You, H., Zhang, Z., Ji, R., & Gao, Y. (2019). Hypergraph Neural Networks. Proceedings of the AAAI Conference on Artificial Intelligence, 33(01), 3558-3565.