Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Hypergraph-enhanced Dual Semi-supervised Graph Classification

Wei Ju, Zhengyang Mao, Siyu Yi, Yifang Qin, Yiyang Gu, Zhiping Xiao, Yifan Wang, Xiao Luo, Ming Zhang

TL;DR

HEAL tackles semi-supervised graph classification under scarce labels by jointly learning higher-order node dependencies via a learnable hypergraph and capturing hyperedge interactions through a line graph. A relational consistency loss facilitates knowledge transfer between the two branches, leveraging unlabeled graphs to regularize representations. The method demonstrates strong performance across six real-world datasets, with ablations confirming the complementary contributions of each component. Overall, HEAL advances graph representation learning by integrating hypergraph and line-graph perspectives to uncover deeper graph semantics and improve classification accuracy.

Abstract

In this paper, we study semi-supervised graph classification, which aims at accurately predicting the categories of graphs in scenarios with limited labeled graphs and abundant unlabeled graphs. Despite the promising capability of graph neural networks (GNNs), they typically require a large number of costly labeled graphs, while a wealth of unlabeled graphs fail to be effectively utilized. Moreover, GNNs are inherently limited to encoding local neighborhood information using message-passing mechanisms, thus lacking the ability to model higher-order dependencies among nodes. To tackle these challenges, we propose a Hypergraph-Enhanced DuAL framework named HEAL for semi-supervised graph classification, which captures graph semantics from the perspective of the hypergraph and the line graph, respectively. Specifically, to better explore the higher-order relationships among nodes, we design a hypergraph structure learning to adaptively learn complex node dependencies beyond pairwise relations. Meanwhile, based on the learned hypergraph, we introduce a line graph to capture the interaction between hyperedges, thereby better mining the underlying semantic structures. Finally, we develop a relational consistency learning to facilitate knowledge transfer between the two branches and provide better mutual guidance. Extensive experiments on real-world graph datasets verify the effectiveness of the proposed method against existing state-of-the-art methods.

Hypergraph-enhanced Dual Semi-supervised Graph Classification

TL;DR

HEAL tackles semi-supervised graph classification under scarce labels by jointly learning higher-order node dependencies via a learnable hypergraph and capturing hyperedge interactions through a line graph. A relational consistency loss facilitates knowledge transfer between the two branches, leveraging unlabeled graphs to regularize representations. The method demonstrates strong performance across six real-world datasets, with ablations confirming the complementary contributions of each component. Overall, HEAL advances graph representation learning by integrating hypergraph and line-graph perspectives to uncover deeper graph semantics and improve classification accuracy.

Abstract

In this paper, we study semi-supervised graph classification, which aims at accurately predicting the categories of graphs in scenarios with limited labeled graphs and abundant unlabeled graphs. Despite the promising capability of graph neural networks (GNNs), they typically require a large number of costly labeled graphs, while a wealth of unlabeled graphs fail to be effectively utilized. Moreover, GNNs are inherently limited to encoding local neighborhood information using message-passing mechanisms, thus lacking the ability to model higher-order dependencies among nodes. To tackle these challenges, we propose a Hypergraph-Enhanced DuAL framework named HEAL for semi-supervised graph classification, which captures graph semantics from the perspective of the hypergraph and the line graph, respectively. Specifically, to better explore the higher-order relationships among nodes, we design a hypergraph structure learning to adaptively learn complex node dependencies beyond pairwise relations. Meanwhile, based on the learned hypergraph, we introduce a line graph to capture the interaction between hyperedges, thereby better mining the underlying semantic structures. Finally, we develop a relational consistency learning to facilitate knowledge transfer between the two branches and provide better mutual guidance. Extensive experiments on real-world graph datasets verify the effectiveness of the proposed method against existing state-of-the-art methods.
Paper Structure (16 sections, 12 equations, 5 figures, 2 tables, 1 algorithm)

This paper contains 16 sections, 12 equations, 5 figures, 2 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Problem Definition & Preliminaries
  3. Methodology
  4. Hypergraph High-order Dependency Learning
  5. Graph Convolution on the Line Graph
  6. Relational Consistency Learning
  7. Computational Complexity Analysis
  8. Experiment
  9. Experimental Setups
  10. Results and Analysis
  11. Ablation Study
  12. Hyper-parameter Study
  13. Analysis of Consistency Loss
  14. Visualization Analysis
  15. Related Work
  16. ...and 1 more sections

Figures (5)

  • Figure 1: Illustration of the proposed framework HEAL.
  • Figure 2: Results of HEAL and baselines with different labeling ratios on PROTEINS and REDDIT-M-5k datasets.
  • Figure 3: Hyper-parameter sensitivity study of HEAL on PROTEINS and COLLAB datasets.
  • Figure 4: Performance comparison with different labeling ratios w.r.t. different types of $\mathcal{L}_{con}$.
  • Figure 5: Visualization of the original graph, hypergraph structure and line graph learned by HEAL (only demonstrating the most significant hyperedges for simplicity).