Source code for the paper VilLain: Self-Supervised Learning on Homogeneous Hypergraphs without Features via Virtual Label Propagation, Geon Lee, Soo Yong Lee, and Kijung Shin, WWW 2024.

Group interactions arise in various scenarios in real-world systems: collaborations of researchers, co-purchases of products, and discussions in online Q&A sites, to name a few. Such higher-order relations are naturally modeled as hypergraphs, which consist of hyperedges (i.e., any-sized subsets of nodes). For hypergraphs, the challenge to learn node representation when features or labels are not available is imminent, given that (a) most real-world hypergraphs are not equipped with external features while (b) most existing approaches for hypergraph learning resort to additional information. Thus, in this work, we propose VilLain, a novel self-supervised hypergraph representation learning method based on the propagation of virtual labels (v-labels). Specifically, we learn for each node a sparse probability distribution over v-labels as its feature vector, and we propagate the vectors to construct the final node embeddings. Inspired by higher-order label homogeneity, which we discover in real-world hypergraphs, we design novel self-supervised loss functions for the v-labels to reproduce the higher-order structure-label pattern. We demonstrate that VilLain is: (a) Requirement-free: learning node embeddings without relying on node labels and features, (b) Versatile: giving embeddings that are not specialized to specific tasks but generalizable to diverse downstream tasks, and (c) Accurate: more accurate than its competitors for node classification, hyperedge prediction, node clustering, and node retrieval tasks.

• Datasets used in the paper (Amazon, Trivago, Cora, Citeseer, Pubmed, DBLP, Primary, and High) are in data.
• The input file is the hypergraph which is in the form:

A pair (V_idx, E_idx) of PyTorch tensors lengthed (# nonzero elements of the incidence matrix)
Node V_idx[i] is contained in hyperedge E_idx[i].

e.g.,
[tensor([     0,      1,      2,  ...,  80632,   5121, 260208]),
 tensor([    0,     0,     0,  ..., 31963, 31963, 31963])]

• The hypergraph data can be loaded using the following command:

with open(os.path.join(data_path, f'H.pickle'), 'rb') as f:
    H = pkl.load(f)

V_idx = H[0]
E_idx = H[1]

• Some main hyperparameters of VilLain are:

dim:            embedding dimension (default: 128)
num_step:       number of propagation steps for training, i.e., k in the paper (default: 4)
num_step_gen:   number of propagation steps for inference, i.e., k' in the paper (default: 10 or 100)
nl:             number of v-labels in each subspace, i.e., d/D (default: 2)

• To run with specified hyperparameters, execute main.py:

python main.py --gpu [GPU ID] --dataset [DATASET NAME] --num_step [# PROPAGATION STEP (TRAINING)] --num_step_gen [# PROPAGATION STEP (INFERENCE)] --lr [LEARNING RATE] --num_labels [NUMBER OF LABELS] --dim [DIMENSION]

e.g.,
python main.py --gpu 0 --dataset cora --num_step 4 --num_step_gen 100 --lr 0.01 --num_labels 2 --dim 128

• To aggregate the embeddings learned with different numbers of v-labels, execute emb_concat.py:

python emb_concat.py --dataset [DATASET NAME] --num_step [# PROPAGATION STEP (TRAINING)] --num_step_gen [# PROPAGATION STEP (INFERENCE)] --dim [DIMENSION]

e.g.,
python emb_concat.py --dataset cora --num_step 4 --num_step_gen 100 --dim 128

• To run a demo, execute the following command in code:

./run.sh

• Embeddings will be saved in code/embs as a PyTorch tensor. The final embedding is saved as _merged.pkl.