From Representation to Clusters: A Contrastive Learning Approach for Attributed Hypergraph Clustering

Li Ni; Shuaikang Zeng; Lin Mu; Longlong Lin

cs.LGMar 10, 2026

From Representation to Clusters: A Contrastive Learning Approach for Attributed Hypergraph Clustering

Authors: Li Ni, Shuaikang Zeng, Lin Mu, Longlong Lin

Organizations: School of Computer Science and Technology, Anhui University, Hefei, China · School of Computer Science and Technology, Southwest University, Chongqing, China

Abstract

Contrastive learning has demonstrated strong performance in attributed hypergraph clustering. Typically, existing methods based on contrastive learning first learn node embeddings and then apply clustering algorithms, such as k-means, to these embeddings to obtain the clustering results.However, these methods lack direct clustering supervision, risking the inclusion of clustering-irrelevant information in the learned graph.To this end, we propose a Contrastive learning approach for Attributed Hypergraph Clustering (CAHC), an end-to-end method that simultaneously learns node embeddings and obtains clustering results. CAHC consists of two main steps: representation learning and cluster assignment learning. The former employs a novel contrastive learning approach that incorporates both node-level and hyperedge-level objectives to generate node embeddings.The latter joint embedding and clustering optimization to refine these embeddings by clustering-oriented guidance and obtains clustering results simultaneously.Extensive experimental results demonstrate that CAHC outperforms baselines on eight datasets.

Figures & tables

Panel of Figure 1 — Figure 1. Comparison of contrastive learning-based hypergraph clustering methods and the proposed CAHC.

Figure 2. Overview of CAHC. CAHC consists of two core steps: hypergraph representation learning and cluster assignment learning. Specifically, hypergraph representation learning optimizes $\mathcal{L}_{\text{hyper}}+\mathcal{L}_{\text{node}}$ to learn node embeddings from the inherent structural and attribute information of the data. Cluster assignment learning optimizes $\mathcal{L}_{\text{hyper}}$ + $\mathcal{L}_{\text{node}}$ + $\mathcal{L}_{\text{clu}}$ to achieve representations and clustering results. A flowchart illustrating the two stages of the CAHC model. Stage 1 shows data augmentation and encoding. Stage 2 shows cluster assignment learning.

Dataset	# Nodes	# Hyperedges	# Features	# Classes
Cora-C	1,434	1,579	1,433	7
Citeseer	1,458	1,079	3,703	6
Pubmed	3,840	7,963	500	3
Cora-A	2,388	1,072	1,433	7
DBLP	41,302	22,363	1425	6
20NewsW100	16,242	100	100	4

Table 1. Statistics of the datasets.

Dataset	Metric	Node2vec	DGI	RAGC	Hyper2vec	TriCL	SE-HSSL	CAHC
Cora-C	ACC	47.4 $\pm$ 5.4	68.4 $\pm$ 3.4	68.6 $\pm$ 1.6	59.7 $\pm$ 5.1	70.0 $\pm$ 1.1	72.0 $\pm$ 3.5	71.4 $\pm$ 1.7
	F1	44.7 $\pm$ 7.1	62.0 $\pm$ 3.0	64.8 $\pm$ 1.6	58.5 $\pm$ 0.4	65.1 $\pm$ 1.8	67.0 $\pm$ 4.3	69.1 $\pm$ 2.8
	NMI	32.5 $\pm$ 3.9	54.0 $\pm$ 1.0	50.9 $\pm$ 1.7	40.6 $\pm$ 1.1	55.2 $\pm$ 0.8	56.5 $\pm$ 2.4	56.4 $\pm$ 0.9
	ARI	20.2 $\pm$ 4.2	47.1 $\pm$ 2.0	43.5 $\pm$ 2.2	33.6 $\pm$ 0.9	49.2 $\pm$ 0.2	49.3 $\pm$ 3.5	50.8 $\pm$ 1.6
Citeseer	ACC	36.2 $\pm$ 1.6	63.8 $\pm$ 4.3	69.4 $\pm$ 1.4	42.7 $\pm$ 1.3	67.6 $\pm$ 1.5	68.2 $\pm$ 1.7	69.8 $\pm$ 0.8
	F1	29.2 $\pm$ 3.1	55.7 $\pm$ 3.4	61.9 $\pm$ 1.9	41.1 $\pm$ 1.2	62.1 $\pm$ 2.0	62.5 $\pm$ 1.5	65.1 $\pm$ 0.6

Table 2. Comparison of clustering results (%) on datasets. All results are reported as mean ± std over 5 independent runs. The best results are in bold . ‘OOM’ denotes the out-of-memory failure, while ‘OOT’ signifies no results are obtained within 24 hours.

Panel of Figure 3 — Figure 3. Impact of clustering guidance.

Dataset	Metric	w/o re	w/o hy	w/o no	w/o cl	w/ mu	CAHC
Cora-C	ACC	28.0	68.8	72.9	70.4	71.6	71.4
	F1	6.2	63.4	70.5	68.4	67.1	69.1
	NMI	0.0	51.4	54.4	55.2	54.5	56.4
	ARI	0.0	46.5	51.9	49.2	49.9	50.8
Citeseer	ACC	21.7	69.1	65.5	68.8	67.0	69.8
	F1	5.9	64.9	60.6	64.3	62.4	65.1

Table 3. Ablation study results (%).

Panel of Figure 4 — Figure 4. Sensitivity analysis of $p_{\text{f}}$ and $p_{\text{m}}$ .

Figure 5. Sensitivity analysis of

D

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

Dataset	$p_{f}$	$p_{m}$	D	LR	Ass. LR	WD	Pre. E	Ass. E
Cora-C	0.4	0.4	512	5e-4	5e-4	5e-4	180	120
Citeseer	0.4	0.4	2048	1e-4	1e-5	5e-5	200	80
Pubmed	0.1	0.4	1024	5e-5	5e-5	5e-5	150	60
Cora-A	0.3	0.2	768	1e-4	1e-5	1e-5	100	140
DBLP	0.4	0.2	256	8e-4	8e-4	5e-5	260	100
20NewsW100	0.1	0.2	512	1e-3	1e-4	5e-4	80	30

Appendix

Table 4. Key hyperparameter settings for our model (CAHC). Fixed parameters are detailed in the text.

Figure 6. Scalability of CAHC on the DBLP, cora-C, Pubmed, and Cora-A datasets.

Panel of Figure 7 — Figure 7. Parameter analysis of the number of heads and attention dimension on the Citeseer and Cora-A datasets. The color intensity corresponds to the NMI score.

Panel of Figure 8 — Figure 8. t-SNE visualization of learned node embeddings on the Cora-A (top row) and 20newW100 (bottom row) datasets. The visualization compares the clustering quality of our proposed model against baselines. Points represent nodes, colored by their ground-truth labels.

Hypergraph Learning

Hypergraph learning models relationships among groups of entities to capture interactions that pairwise graphs cannot represent.

See all 54 papers

HyperFuse: Fast Self-Supervised Node Embeddings for Attributed Hypergraphs
Oct 2, 2026Megha P, Harshit Kumar, Srajan Agarwal +3Hypergraph Learning Hypergraph Neural Networks
School of Data Science, Indian Institute of Science Education and Research Thiruvananthapuram, Thiruvananthapuram, 695551, Kerala, India. · Department of Mathematics and Statistics, Indian Institute of Science Education and Research Kolkata, Mohanpur, 741246, West Bengal, India. · Institute of Computer Science, University of Rostock, Rostock, 18051, Germany. +1
Bridge the Gaps: Heterogeneous Attributed Graph Clustering via Quaternion Representation Learning
Jun 22, 2026Xinxi Chen, Junyang Chen, Yiqun Zhang +2Graph Structure Learning Graph Neural Networks
School of Computer Science and Technology, Guangdong University of Technology, Guangzhou 510006, China
Robust Contrastive Graph Clustering with Adaptive Local-Global Integration
May 27, 2026Lei Zhang, Fubo Sun, Haipeng Yang +2Graph Representation Learning Multi-Scale Feature Fusion
School of Computer Science and Technology, Anhui University · College of Management and Economics, Tianjin University

Oct 2, 2026J/K move · Enter open · S save

HyperFuse: Fast Self-Supervised Node Embeddings for Attributed Hypergraphs

Megha P, Harshit Kumar, Srajan Agarwal, Anirban Banerjee+2

School of Data Science, Indian Institute of Science Education and Research Thiruvananthapuram, Thiruvananthapuram, 695551, Kerala, India. · Department of Mathematics and Statistics, Indian Institute of Science Education and Research Kolkata, Mohanpur, 741246, West Bengal, India. · Institute of Computer Science, University of Rostock, Rostock, 18051, Germany.+1

Hypergraph Learning Hypergraph Neural Networks

Open paper AskPDF arXiv

Self-supervised hypergraph representation learning can produce informative node embeddings, but existing methods often require deep encoders trained for hundreds of epochs, making embedding generation costly even for hypergraphs with a few thousand nodes. This limits applications requiring embeddings for many or evolving hypergraphs. We present HyperFuse, a label-free pipeline for fast hypergraph representation learning. HyperFuse (i) computes structural node coordinates by maximizing a spectral relaxation of hypergraph modularity using Banerjee's hypergraph adjacency and a matrix-free operator with cost linear in node-hyperedge incidences; (ii) constructs multi-scale feature summaries and assigns bounded utility weights to hyperedges based on member stability under feature and membership masking; and (iii) trains a lightweight utility-weighted hypergraph encoder for 100 epochs using an invariance-decorrelation objective. We compare HyperFuse with TriCL, SE-HSSL, VilLain, and HypeBoy on nine public hypergraphs using six downstream classifiers and k-means clustering. On the eight datasets where all methods completed, HyperFuse required 8.7 s per dataset on average, achieving 13-179x geometric-mean speed-ups over the baselines. It achieved the highest average accuracy with five of six classifiers, while classification and clustering performance was not significantly different from TriCL and SE-HSSL. Compared with HypeBoy, HyperFuse was 13x faster and 2.1-4.1 percentage points more accurate across all classifiers. HyperFuse provides a practical approach for fast, repeated hypergraph embedding generation.

Abstract

Hypergraph Learning

Explore similar work

HyperFuse: Fast Self-Supervised Node Embeddings for Attributed Hypergraphs

Bridge the Gaps: Heterogeneous Attributed Graph Clustering via Quaternion Representation Learning

Robust Contrastive Graph Clustering with Adaptive Local-Global Integration