Closer through commonality: Enhancing hypergraph contrastive learning with shared groups

TL;DR

HyFi tackles exploiting high-order information in hypergraphs for contrastive learning by introducing weak positive pairs and topology-preserving augmentation through node feature perturbation. It couples a HGNN encoder with dual projection heads to produce node and hyperedge embeddings and defines node- and edge-level contrastive losses guided by shared hyperedges, forming a final objective $\mathcal{L}=\sum_i \mathcal{L}_n(z_i,z_m') + \alpha \sum_i \mathcal{L}_e(y_i,y_m')$. The approach demonstrates superior node classification performance across 10 datasets while being more efficient in training time and memory than existing HGCL methods, underscoring the importance of commonality in hypergraphs and offering a scalable path for hypergraph representation learning. The work provides practical, augmentation-friendly techniques that preserve topology and leverage high-order interactions, enabling broad applicability across domains.

Abstract

Hypergraphs provide a superior modeling framework for representing complex multidimensional relationships in the context of real-world interactions that often occur in groups, overcoming the limitations of traditional homogeneous graphs. However, there have been few studies on hypergraphbased contrastive learning, and existing graph-based contrastive learning methods have not been able to fully exploit the highorder correlation information in hypergraphs. Here, we propose a Hypergraph Fine-grained contrastive learning (HyFi) method designed to exploit the complex high-dimensional information inherent in hypergraphs. While avoiding traditional graph augmentation methods that corrupt the hypergraph topology, the proposed method provides a simple and efficient learning augmentation function by adding noise to node features. Furthermore, we expands beyond the traditional dichotomous relationship between positive and negative samples in contrastive learning by introducing a new relationship of weak positives. It demonstrates the importance of fine-graining positive samples in contrastive learning. Therefore, HyFi is able to produce highquality embeddings, and outperforms both supervised and unsupervised baselines in average rank on node classification across 10 datasets. Our approach effectively exploits high-dimensional hypergraph information, shows significant improvement over existing graph-based contrastive learning methods, and is efficient in terms of training speed and GPU memory cost. The source code is available at https://github.com/Noverse0/HyFi.git.

Figures (5)

• Figure 1: Example of tranditional hypergraph contrastive learning vs. HyFi. In (a), contrastive learning only learns postive or negative pairs between graphs augmented by randomly dropping hyperedges or nodes. In (b), HyFi preserves the topology of the hypergraph by adding noise to the node features, while adding weak postive pair relationships for fine-grained contrastive learning.
• Figure 2: The average of the cosine similarity of the node features over the number of commonalities between two nodes, where commonality, defined as a common group, means common hyperedges. The higher the number of common hyperedges, the higher the similarity of the two node features.
• Figure 3: Illustration of the node-level contrastive loss calculation process in HyFi. HyFi uses both the origin view and the noise view to calculate contrastive loss. In the origin view, there is a negative and a weak positive pair, and in the noise view there is a positive pair, which is a node in the same position as the anchor node.
• Figure 4: Illustration of the GPU memory usage comparison between HyFi and TriCL, focusing on how they differ in memory consumption in various dimensions, where dimension denotes the encoder output dimension and the project head output dimension, a bar with a diagonal pattern means 'Out of Memory'.
• Figure 5: Illustration of the performance of node classification as the number of positive samples. The number of noise views is equal to the number of positive samples in the contrastive learning.