Adaptive Temperature Enhanced Dual-level Hypergraph Contrastive Learning

Yiyue Qian, Tianyi Ma, Chuxu Zhang, Yanfang Ye

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Keywords: hypergraph, contrastive learning, self-supervised learning.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

TL;DR: Explore the group-wise behaviors and the temperature index in hypergraph contrastive learning.

Abstract:

Hypergraphs, which incorporate hyperedges to link multiple nodes and capture complex high-order relationships, have attracted increasing attention in recent years.
Consequently, a bunch of hypergraph neural networks has been proposed to model the high-order relationships between hyperedges and nodes. Inspired by the success of graph contrastive learning, researchers have begun exploring the benefits of contrastive learning over hypergraphs. However, these works still have the following limitations in modeling the high-order relationships over unlabeled data: (i) They primarily focus on maximizing the agreements among individual node embeddings while neglecting the capture of group-wise collective behaviors within hypergraphs; (ii) Most of them disregard the importance of the temperature index in discriminating contrastive pairs during contrast optimization. To address these limitations, we propose a novel \textbf{Ad}aptive \textbf{T}emperature enhanced \textbf{Hy}per\textbf{G}raph \textbf{C}ontrastive \textbf{L}earning framework called \textbf{AdT-HyGCL} to boost contrastive learning over hypergraphs. Specifically, we first introduce a noise enhancement module to generate relatively challenging augmented hypergraphs for hypergraph contrastive tasks. Unlike most works that merely maximize the agreement of node embeddings in hypergraphs, we then propose a dual-level contrast mechanism that not only captures the individual node behaviors in a local context but also models the group-wise collective behaviors of nodes within hyperedges from a community perspective. Furthermore, we design an adaptive temperature-enhanced contrastive optimization to improve the discrimination ability between positive and negative contrastive pairs, thereby facilitating more effective hypergraph representation learning. Theoretical justifications and empirical experiments conducted on eight benchmark hypergraphs demonstrate that AdT-HyGCL exhibits excellent rationality, generalization, effectiveness, and robustness compared to state-of-the-art baseline models.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 6979