[ICLR 2024] HypeBoy: Generative Self-Supervised Representation Learning on Hypergraphs

Authors: Sunwoo Kim, Shinhwan Kang, Fanchen Bu, Soo Yong Lee, Jaemin Yoo, and Kijung Shin.

Paper: https://openreview.net/forum?id=DZUzOKE6og

Published as a conference paper at ICLR 2024

Dataset Description

We provide 11 hypergraph benchmark datasets that are used in our work. Dataset statistics are as follows:

Dataset	# of nodes	# of hyperedges	# of features	# of classes
Cora	1,434	1,579	1,433	7
Citeseer	1,458	1,079	3,703	6
Pubmed	3,840	7,963	500	3
Cora-CA	2,388	1,072	1,433	7
DBLP-P	41,302	22,263	1,425	6
DBLP-A	2,591	2,690	334	4
AMiner	20,201	8,052	500	12
IMDB	3,939	2,015	3066	3
ModelNet-40	12,311	12,311	100	40
News	16,242	100	4	4
House	1,290	341	100	2

(Due to the file size of datasets) Datasets can be found at https://www.dropbox.com/scl/fi/pes0jgk0syrxosk71n1ua/datasets.zip?rlkey=h1r7j745z1a73qlrd6pv9ypcd&dl=0.

Details of each dataset file are provided at README.txt in the provided datasets.zip of the above link.

Code Description

Overview

We provide code for reproducing HypeBoy's performance at:

• Fine-tuning of node classification,
• Linear evaluation of node classification,
• Linear evaluation of hyperedge prediction.

Pre-requisite

The overall hierarchy should be set as follows:

/datasets
  |
  /cora_cite
    |
    X.pt
    H.pt
    Y.pt
    edge_bucket.pickle
    data_split_0.01.pickle
  /citeseer_cite
    |...
  ...
main.py
src.py
HNNs.py

Run code

The crucial command is as follows:

python3 main.py -data "str: Data name" -task "str: Downstream task" -epoch "int: SSL epoch" -p_x "float: $p_{x}$" -p_e "float: $p_{e}$" -device "str: Cuda device"

Example:

python3 main.py -data cora_cite -task finetuning -epoch 200 -p_x 0.4 -p_e 0.8 -device cuda:0

Details of each component are as follows:

-data (String): Name of datasets you want to use. It should be given one of

• cora_cite: Cora dataset
• citeseer_cite: Citeseer dataset
• pubmed_cite: Pubmed dataset
• cora_ca: Cora-Coauthorship dataset
• dblp_coauth: DBLP-P dataset
• dblp_copub: DBLP-A dataset
• aminer: AMiner dataset
• imdb: IMDB dataset
• modelnet_40: ModeltNet 40 dataset
• news: News dataset
• house: House dataset

E.g., -data cora_cite

-task (String): The task you want to evaluate. It should be given one of

• finetuning: Fintuning protocol of node classification
• linear_node: Linear evaluation protocol of node classification
• linear_edge: Linear evaluation protocol of hyperedge prediction

E.g., -task finetuning

-epoch (Integer): Number of epochs you want to run HypeBoy. It should be given as an integer.

E.g., -epoch 200

-p_x (Float): Node feature drop probability $p_{x}$. Note that $p_{x} \in [0, 1]$ should hold.

E.g., -p_x 0.4

-p_x (Float): Hyperedge drop probability $p_{e}$. Note that $p_{e} \in [0, 1]$ should hold.

E.g., -p_x 0.5

-device (String): Cuda device machine name.

E.g., -device cuda:0