《计算机应用》唯一官方网站 ›› 2026, Vol. 46 ›› Issue (4): 1042-1049.DOI: 10.11772/j.issn.1001-9081.2025050540

• 人工智能 • 上一篇    下一篇

基于块对角表示的超图学习方法

张生伟1, 王豪2, 金泰松2()   

  1. 1.中国航空工业集团公司洛阳电光设备研究所,河南 洛阳 471000
    2.厦门大学 信息学院,福建 厦门 361102
  • 收稿日期:2025-05-19 修回日期:2025-08-18 接受日期:2025-08-27 发布日期:2025-08-28 出版日期:2026-04-10
  • 通讯作者: 金泰松
  • 作者简介:张生伟(1982—),男,河南南阳人,研究员,硕士,主要研究方向:计算机视觉、模式识别
    王豪(2001—),男,河南安阳人,硕士研究生,主要研究方向:计算机视觉、模式识别
  • 基金资助:
    国家自然科学基金面上项目(62072386);福建省自然科学基金资助项目(2025J01003)

Hypergraph learning method via block diagonal representation

Shengwei ZHANG1, Hao WANG2, Taisong JIN2()   

  1. 1.Luoyang Institute of Electro-Optical Equipment of AVIC,Luoyang Henan 471000,China
    2.School of Informatics,Xiamen University,Xiamen Fujian 361102,China
  • Received:2025-05-19 Revised:2025-08-18 Accepted:2025-08-27 Online:2025-08-28 Published:2026-04-10
  • Contact: Taisong JIN
  • About author:ZHANG Shengwei, born in 1982, M. S., research fellow. His research interests include computer vision, pattern recognition.
    WANG Hao, born in 2001, M. S. candidate. His research interests include computer vision, pattern recognition.
  • Supported by:
    General Program of National Natural Science Foundation of China(62072386);Fujian Provincial Natural Science Foundation(2025J01003)

摘要:

作为一种能够自然表征多元数据对象间高阶关系的数学工具,超图相较于传统图机器学习方法展现出显著优势。基于超图的机器学习范式的前提在于通过超图学习方法构建能够反映数据间高阶关系的超图。然而,现有的超图学习方法在应对噪声和数据损坏方面的鲁棒性不足制约了它们的实际应用效果。为了解决这个问题,提出一种基于块对角表示的超图学习方法。该方法优化一个引入块对角约束的数据重构目标函数,并利用获得的重构系数生成超边和设置超边权重。在加入噪声的图像数据集上的实验结果表明:与CR-HG(CorrentRopy-induced low-rank HyperGraph)方法相比,所提方法在加入高斯噪声的噪声率为40%和加入椒盐噪声的噪声密度为30%的Coil20图像集上的互信息(NMI)分别提升了2.6和1.0个百分点;在加入高斯噪声的噪声率40%和加入椒盐噪声的噪声密度为30%的USPS图像集上的分类准确率(ACC)分别提升了2.1和1.1个百分点。可见,所提方法的学习性能优于现有的主流超图学习方法。

关键词: 块正交, 超边生成, 权重设置, 超图, 高阶关系

Abstract:

As a mathematical tool of representing the high-order relationship among multiple data objectives naturally, hypergraphs exhibit significant advantages compared to traditional graph-based machine learning methods. The premise of hypergraph-based learning paradigm lies in constructing hypergraph that can reflect the high-order relationships via hypergraph learning methods. However, the lack of robustness of the existing hypergraph learning methods in dealing with noise and data corruption limits their real application effects. To address this issue, a hypergraph learning method via block diagonal representation was proposed. In the method, an objective function of data reconstruction introducing the block diagonal constraint was optimized, and hyperedges were generated and the hyperedge weights were set using the obtained reconstruction coefficients. Experimental results on two image datasets adding noise demonstrate that compared with CR-HG (CorrentRopy-induced low-rank HyperGraph) method, the proposed method achieves improvements of 2.6 and 1.0 percentage points in Normalized Mutual Information (NMI) on the Coil20 image set with 40% Gaussian noise ratio and 30% salt-and-pepper noise density, respectively. Additionally, on the USPS image set with 40% Gaussian noise ratio and 30% salt-and-pepper noise density, the proposed method increases the Classification Accuracy for Classification(ACC) by 2.1 and 1.1 percentage points, respectively. It can be seen that the learning performance of the proposed method is superior to that of the existing mainstream hypergraph learning methods.

Key words: block orthogonality, hyperedge generation, weight setting, hypergraph, high-order relationship

中图分类号: