Official code implementation of ZEN

This repository is the official implementation of Parameter-Free Hypergraph Neural Network for Few-Shot Node Classification.

Dataset description

We provide ten hypergraph benchmark datasets for evaluating accuracy, and one additional dataset for interpretability analysis. Their overall statistics are as follows:

Benchmark datasets

Name	Command name	# of nodes	# of edges	# of classes	# of features
Cora	cora	2,708	1,579	7	1,433
Citeseer	citeseer	3,312	1,079	6	3,703
Pubmed	pubmed	19,717	7,963	3	500
Cora-CA	coauthor_cora	2,708	1,072	7	1,433
20News	20newsW100	16,242	100	4	100
MN40	ModelNet40	12,311	12,311	40	100
Congress	congress-bills	1,718	83,105	2	100
Walmart	walmart-trips	88,860	69,906	11	100
Senate	senate-committees	282	315	2	100
House	house-committees	1,290	340	2	100

Interpretability dataset

Name	Command name	# of nodes	# of edges	# of classes	# of features
Zoo	zoo	101	43	7	16

Evaluation

To evaluate ZEN on Cora, run:

python main.py

Additional arguments

Argument	Description
-data	Name of the dataset to use. Options: cora, citeseer, pubmed, etc. Default: cora.
-n	Grid resolution for 2-simplex search during hyperparameter tuning. Total grid points: $(n+2)(n+1)/2$. Default: 9.
-k	Number of labeled nodes per class used for training and validation in few-shot classification. Default: 5.
-run	Number of independent random splits (train/val/test) for evaluation. Default: 10.
-device	Computation device. Options: cuda:0, cuda:1, cpu, etc. Default: cuda:0.