• Title: Datasets, Tasks, and Training Methods for Benchmarking Large-scale Hypergraph Learning.
• Authors: Sunwoo Kim, Dongjin Lee, Yul Kim, Jungho Lee, Taeho Hwang, and Kijung Shin.
• Venue: Data Mining and Knowledge Discovery 2023. (+ Journal Track of ECML-PKDD 2023)

We provide hypergraph datasets at the below links:

• Static and Temporal (Unzip): https://www.dropbox.com/sh/qhz6rqol5mue4wc/AAByciJuLNba8uGv8MdMpbf-a?dl=0.
• Static (Zip): https://www.dropbox.com/sh/1908u1h5w6zxzjs/AABP04dM4VU_0fadrpe_5pb-a?dl=0.

We provide

• Each dataset's feature, node label, original hyperedge information, split hyperedge information (for task 1)
• Each dataset's partitioned hypergraph (# of Partition $|P|$ / DBLP:4, Trivago:32, OGBN_MAG:128, AMINER and MAG: 256)
• For DBLP, Trivago, OGBN_MAG, we also provide partitioned hypergraph P-IOS partition.

Refer to README_DATA.txt file for more details regarding datasets.

In this repository, we provide source codes for

• Obtaining performance on proposed task 1 (hyperedge disambiguation)
• Obtaining performance on proposed task 2 (local clustering)

For dataset one aims to run the code, files of the corresponding dataset in the above links should be located in scr/data folder. For example

src
  |_ data
      |_ aminer
            |_ aminer_X_vec.pt
            |_ aminer_y.pt
            |_ aminer_E.pt
            |_ ...

We provide a hyperparameter combination for the reproducibility of experimental results.
Refer to the best_hyperparameter directory, where we saved each dataset-model combination's best hyperparameter as .json files.

One can simply run code with codes in src folder as follows;
For task 1 (hyperedge disambiguation)

python experiment1 -data "data" -model "learning-method" -device "GPU-device" -lr 0.001 -seed 0 

For task 2 (local clustering)

python experiment2 -data "data" -model "learning-method" -device "GPU-device" -lr 0.001 -seed 0 

For contrastive learning, additional arguments can be given as

python experiment1 -data "data" -model "learning-method" -device "GPU-device" -lr 0.001 -seed 0 -n_neg 1 -d_rate 0.3 -ep 25

Arguments are

• -data : Dataset one wants to perform experiments. (Possible Options: "dblp", "trivago", "ogbn_mag", "aminer", "mag")
• -model : Training Method one wants to perform experiments. (Possible Options: "mlp", "full_ssl", "full_cl", "part_ssl", "HCL", "HCL_PINS", "HCL_PIOS")
• -device : GPU machine (e.g. "cuda:0")
• -lr : Learning rate for method. (float / e.g. 0.001)
• -n_neg : Applicable to CL methods. Number of negative samples used for CL training. (int / e.g. 2)
• -d_rate : Applicable to CL methods. Feature & Incidence matrix dropping probability. (float / e.g. 0.3)
• -ep : Applicable to CL methods. Contrastive encoder's checkpoint for contrastive training. (int / e.g. 25)
• -seed : Dataset & model initialization seed. (int / e.g. 0)