基于分数阶图神经扩散的跨频域对齐对比学习方法

王友清; 徐世龙; 赵天祥; 王宇晨; 辛梦媛; 张琦; 苏烨; 郭继鹏

doi:10.16383/j.aas.c250604

基于分数阶图神经扩散的跨频域对齐对比学习方法

doi: 10.16383/j.aas.c250604 cstr: 32138.14.j.aas.c250604

王友清^1,,
徐世龙^1,,
赵天祥^1,,
王宇晨^1,,
辛梦媛^1,,
张琦^2,,
苏烨^3,,
郭继鹏^1,

1.
北京化工大学信息科学与技术学院北京 100029
2.
山东农业大学信息科学与工程学院泰安 271018
3.
中国科学院深圳先进技术研究院深圳 518055

基金项目: 国家重点研发项目(2024YFB3311405), 国家自然科学基金(62225303, 62403043)资助

详细信息

作者简介:
王友清：北京化工大学信息科学与技术学院教授. 主要研究方向为工业大数据分析. E-mail: wang_youqing@buct.edu.cn

徐世龙：北京化工大学信息科学与技术学院硕士研究生. 主要研究方向为图神经网络和图表示学习. E-mail: xushilong@buct.edu.cn

赵天祥：北京化工大学信息科学与技术学院硕士研究生. 主要研究方向为机器学习. E-mail: tianxiangzhao@buct.edu.cn

王宇晨：北京化工大学信息科学与技术学院硕士研究生, 主要研究方向为图神经网络和图异常检测. E-mail: 2025200789@buct.edu.cn

辛梦媛：北京化工大学信息科学与技术学院博士研究生. 主要研究方向为对抗样本学习. E-mail: 2025400300@buct.edu.cn

张琦：山东农业大学信息科学与工程学院副教授. 主要研究方向为图机器学习, 知识图谱和大语言模型. E-mail: QiZhang2025@sdau.edu.cn

苏烨：中国科学院深圳先进技术研究院博士研究生. 主要研究方向为机器学习与模式识别. E-mail: ye.su@siat.ac.cn

郭继鹏：北京化工大学信息科学与技术学院副教授. 主要研究方向为机器学习与模式识别和工业过程数据分析. 本文通讯作者. E-mail: guojipeng@buct.edu.cn

计量
- 文章访问数: 13
- HTML全文浏览量: 5
- 被引次数: 0
出版历程
- 收稿日期: 2025-11-05
- 录用日期: 2026-01-20
- 网络出版日期: 2026-03-30

Fractional-order Graph Neural Diffusion for Cross-frequency Alignment Contrastive Learning

1.
College of Information Science and Technology, Beijing University of Chemical Technology, Beijing 100029
2.
School of Information Science and Engineering, Shandong Agricultural University, Taian 271018
3.
Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences, Shenzhen 518055

Funds: Supported by National Key Research and Development Program of China (2024YFB3311405) and National Natural Science Foundation of China (62225303, 62403043)

More Information

Author Bio:
WANG You-Qing　Professor at the College of Information Science and Technology, Beijing University of Chemical Technology. His main research interest is industrial big data analysis

XU Shi-Long　Master student at the College of Information Science and Technology, Beijing University of Chemical Technology. His research interests include graph nerual network and graph represention learning

ZHAO Tian-Xiang　Master student at the College of Information Science and Technology, Beijing University of Chemical Technology. His main research interest is machine learning

WANG Yu-Chen　Master student at the College of Information Science and Technology, Beijing University of Chemical Technology. His research interests include graph neural networks and graph anomaly detection

XIN Meng-Yuan　Ph.D. candidate at the College of Information Science and Technology, Beijing University of Chemical Technology. Her main research interest is adversarial example learning

Zhang Qi　Associate professor at School of Information Science and Engineering, Shandong Agricultural University. His research interests include graph machine learning, knowledge graph and large language models

SU Ye　Ph.D. candidate at Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences. His research interests include machine learning and pattern recognition

GUO Ji-Peng　Associate professor at the College of Information Science and Technology, Beijing University of Chemical Technology. His research interests include machine learning, pattern recognition and industrial process data analysis. Corresponding author of this paper

摘要

摘要: 图对比学习(GCL)作为一种强大的自监督表示学习范式, 能够通过有效利用无标签数据来增强半监督学习中的表示判别性和泛化能力. 然而, 现有的GCL方法在学习判别性嵌入表示以及图数据增强过程中实现对比多样性与语义不变性之间的平衡方面存在困难, 这导致在构建增强视图时关键信息的丢失. 为了解决这些挑战, 提出了一种新颖的跨频域对齐对比学习(CfACL)框架, 利用分数阶图神经扩散(FGND)进行图节点表示学习. FGND利用切比雪夫多项式分数阶微分方程实现图信号中多阶邻域信息的远程扩散, 缓解过平滑问题并提高图嵌入表示的判别能力. 随后, 通过高频和低频滤波器分别构建两种不同的FGND形式, 形成自然的增强对比视图, 避免了随机增强引起的内在结构坍塌和语义偏移. CfACL方法将高频滤波分量转换到低频域, 并在镜像的虚拟谱空间中进行对比学习, 从而能够在全局一致的语义空间中吸收有益的高频细节, 为下游任务生成全面的表示. 在同配性和异配性基准图数据集上的大量节点分类实验结果验证了所提方法的有效性.
- 图对比学习 /
- 分数阶图扩散 /
- 高频滤波 /
- 低频滤波 /
- 跨频域对齐
Abstract: Graph Contrastive Learning (GCL), a powerful self-supervised representation learning paradigm, could enhance representation discriminability and generalization in semi-supervised learning by effectively leveraging unlabeled data. However, existing GCL methods struggle to learn discriminative embedding and achieve better balance between contrastive diversity and semantic invariance during graph data augmentation, inevitably leading to the critical information loss when constructing augmented views. To address these challenges, this paper proposes a novel Cross-frequency Alignment Contrastive Learning (CfACL) framework with Fractional-order Graph Neural Diffusion (FGND). The FGND leverages Chebyshev polynomial fractional differential equations to achieve the long-range diffusion of multi-order neighboring information, alleviating over-smoothing and improving the discriminability of graph representation. Then, two distinct FGNDs are characterized by high-frequency and low-frequency filters to form natural augmented contrastive views, avoiding the intrinsic structure collapse and semantic shift caused by random augmentation. The CfACL transforms high-frequency components into the low-frequency domain and achieves the contrastive learning in mirrored virtual spectral space, which is capable of absorbing beneficial high-frequency details in a globally consistent semantic space and results in comprehensive representation for downstream task. Extensive node classification experiments demonstrate the effectiveness of the proposed method across homophilic and heterophilic benchmark graph datasets.
- graph contrastive learning /
- fractional-order graph diffusion /
- high-frequency filtering /
- low-frequency filtering /
- cross-frequency alignment

HTML全文

图 1 消融实验结果

Fig. 1 The results of ablation experiment

下载: 全尺寸图片幻灯片

图 2 CfACL使用不同对比学习损失在数据集Cornell上的实验结果

Fig. 2 The experimental results of CfACL with various contrastive learning losses on the Cornell dataset

下载: 全尺寸图片幻灯片

图 3 CfACL关于不同时间t的实验结果

Fig. 3 The experimental results of CfACL with various time t

下载: 全尺寸图片幻灯片

图 4 CfACL在时间$ t\in \{4,\; 8,\; 16,\; 32,\; 64\} $下的分类性能

Fig. 4 The classification performance of CfACL when time $ t\in \{4,\; 8,\; 16,\; 32,\; 64\} $

下载: 全尺寸图片幻灯片

图 5 在Cora和Cornell数据集上的t-SNE可视化结果

Fig. 5 The t-SNE visualization results on the Cora and Cornell datasets

下载: 全尺寸图片幻灯片

图 6 在Cora和Chameleon数据集上的训练损失和分类准确率随训练迭代步数的变化

Fig. 6 The variation of training loss and classification accuracy with training iteration epochs on Cora and Chameleon datasets

下载: 全尺寸图片幻灯片

表 1 数据集统计信息

Table 1 The statistics of the datasets

数据集	节点数量	边数量	特征数量	类别数量	边同配性比
Cora	2708	5429	1433	7	0.81
Citeseer	3327	4732	3703	6	0.73
PubMed	19717	88651	500	3	0.80
Wisconsin	251	466	1703	5	0.17
Texas	183	309	1793	5	0.06
Cornell	183	295	1703	5	0.12
Chameleon	2277	36101	2325	5	0.23
Squirrel	5201	217073	2089	5	0.20
Actor	7600	33391	932	5	0.21

下载: 导出CSV

表 2 在同配和异配数据集上的节点分类实验结果

Table 2 The node classification experiment results on homophilic and heterophilic datasets

方法	Cora	Citeseer	PubMed	Wisconsin	Cornell	Texas	Chameleon	Squirrel	Actor
DGI	$ 82.30\pm0.60 $	$ 71.80\pm0.70 $	$ 76.80\pm0.60 $	$ 55.21\pm1.02 $	$ 45.33\pm6.11 $	$ 58.53\pm2.98 $	$ 60.27\pm0.70 $	$ 26.44\pm1.12 $	$ 28.30\pm0.76 $
GCA	$ 82.93\pm0.42 $	$ 72.19\pm0.31 $	$ 80.79\pm0.45 $	$ 59.55\pm0.81 $	$ 52.31\pm1.09 $	$ 52.92\pm0.46 $	$ 63.66\pm0.32 $	$ 48.09\pm0.21 $	$ 28.47\pm0.29 $
CCA-SSG	$ 84.00\pm0.40 $	$ 73.10\pm0.30 $	$ 81.00\pm0.40 $	$ 58.46\pm0.96 $	$ 52.17\pm1.04 $	$ 59.89\pm0.78 $	$ 62.41\pm0.22 $	$ 46.76\pm0.36 $	$ 27.82\pm0.60 $
BGRL	$ 82.70\pm0.60 $	$ 71.10\pm0.80 $	$ 79.60\pm0.50 $	$ 51.23\pm1.17 $	$ 50.33\pm2.29 $	$ 52.77\pm1.98 $	$ 64.86\pm0.63 $	$ 36.22\pm1.97 $	$ 28.80\pm0.54 $
SP-GCL	$ 83.16\pm0.13 $	$ 71.96\pm0.42 $	$ 79.16\pm0.84 $	$ 60.12\pm0.39 $	$ 52.29\pm1.21 $	$ 59.81\pm1.33 $	$ 65.28\pm0.53 $	$ 52.10\pm0.67 $	$ 28.94\pm0.69 $
GraphACL	$ \underline{84.20\pm0.31} $	$ \underline{73.63\pm0.22} $	$ \underline{82.02\pm0.15} $	$ 69.22\pm0.40 $	$ \underline{59.33\pm1.48} $	$ 71.08\pm0.34 $	$ \underline{69.12\pm0.24} $	$ \underline{54.05\pm0.13} $	$ 30.03\pm0.13 $
PolyGCL	$ 81.97\pm0.19 $	$ 71.97\pm0.29 $	$ 77.48\pm0.39 $	$ \underline{76.08\pm3.33} $	$ 43.78\pm3.51 $	$ \underline{72.16\pm3.51} $	$ 46.84\pm1.53 $	$ 34.25\pm0.66 $	$ \underline{34.37\pm0.69} $
CfACL	$ {\bf{85.17}}\pm{\bf{1.51}} $	$ {\bf{76.67}}\pm{\bf{1.38}} $	$ {\bf{86.37}}\pm{\bf{1.26}} $	$ {\bf{80.34}}\pm{\bf{0.47}} $	$ {\bf{71.70}}\pm{\bf{7.43}} $	$ {\bf{83.14}}\pm{\bf{5.34}} $	$ {\bf{72.29}}\pm{\bf{1.50}} $	$ {\bf{62.54}}\pm{\bf{1.14}} $	$ {\bf{36.76}}\pm{\bf{1.34}} $

下载: 导出CSV

表 3 CfACL与代表性对比方法PolyGCL和GraphACL在不同数据集上的训练成本对比(秒)

Table 3 The comparison of training costs for CfACL versus the representative comparison methods PolyGCL and GraphACL across different datasets (Seconds)

方法	Cora	PubMed	Chameleon	Cornell	Texas	Actor
PolyGCL	23.42	4392.96	19.62	32.29	24.03	54.71
GraphACL	23.08	340.35	212.19	28.10	27.64	3888.55
CfACL	19.58	108.86	19.58	22.79	29.96	32.99

下载: 导出CSV

参考文献(41)

[1]	Liang K, Meng L Y, Liu M, Liu Y, Tu W X, Wang S W, et al. A survey of knowledge graph reasoning on graph types: static, dynamic, and multi-modal. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024, 46(12): 9456−9478 doi: 10.1109/TPAMI.2024.3417451
[2]	Li M R, Zhang Y, Zhang W, Zhao S Y, Piao X L, Yin B C. Csat: Contrastive sampling-aggregating transformer for community detection in attribute-missing networks. IEEE Transactions on Computational Social Systems, 2024, 11(2): 2277−2290 doi: 10.1109/TCSS.2023.3292145
[3]	罗彪, 胡天萌, 周育豪, 黄廷文, 阳春华, 桂卫华. 多智能体强化学习控制与决策研究综述. 自动化学报, 2025, 51(3): 510−539 doi: 10.16383/j.aas.c240392 Luo Biao, Hu Tian-Meng, Zhou Yu-Hao, Huang Ting-Wen, Yang Chun-Hua, Gui Wei-Hua. Survey on multi-agent reinforcement learning for control and decision-making. Acta Automatica Sinica, 2025, 51(3): 510−539 doi: 10.16383/j.aas.c240392
[4]	Wu Z H, Lu J L, Yu J J, Zhou S, Pi Y Y, Wang H S. Divide and conquer: Coordinating multiplex mixture of graph learners to handle multi-omics analysis. In: Proceedings of the Thirty-Fourth International Joint Conference on Artificial Intelligence. Montreal, Canada: IJCAI, 2025. 6615-6623
[5]	Li M R, Zhang P Y, Xing W B, Zheng Y J, Zaporojets K, Chen J Z, et al. A survey of large language models for data challenges in graphs. Expert Systems with Applications. Expert Systems with Applications, 2025, 298: 129643
[6]	吴博, 梁循, 张树森, 徐睿. 图神经网络前沿进展与应用. 计算机学报, 2022, 45(1): 35−68 doi: 10.11897/SP.J.1016.2022.00035 Wu Bo, Liang Xun, Zhang Shu-Sen, Xu Rui. Advances and applications in graph neural network. Chinese Journal of Computers, 2022, 45(1): 35−68 doi: 10.11897/SP.J.1016.2022.00035
[7]	Wu Z H, Cai J Y, Zhang Y H, Lu J L, Chen Z L, Zhuang S M, et al. Where graph meets heterogeneity: Multi-view collaborative graph experts. In: Proceedings of 39th Annual Conference on Neural Information Processing Systems. 2025.
[8]	Wu Z H, Pan S R, Chen F W, Long G D, Zhang C Q, Philip S Y. A comprehensive survey on graph neural networks. IEEE Transactions on Neural Networks and Learning Systems, 2021, 32(1): 4−24 doi: 10.1109/TNNLS.2020.2978386
[9]	张建朋, 裴雨龙, 刘聪, 李邵梅, 陈鸿昶. 基于因子图模型的动态图半监督聚类算法. 自动化学报, 2020, 46(4): 670−680 doi: 10.16383/j.aas.c170363 Zhang Jian-Peng, Pei Yu-Long, Liu Cong, Li Shao-Mei, Chen Hong-Chang. A semi-supervised clustering algorithm based on factor graph model for dynamic graphs. Acta Automatica Sinica, 2020, 46(4): 670−680 doi: 10.16383/j.aas.c170363
[10]	Cai L, Li J D, Wang J, Ji S W. Line graph neural networks for link prediction. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022, 44(9): 5103−5113 doi: 10.1007/978-981-16-6054-2_10
[11]	公沛良, 艾丽华. 用于半监督分类的二阶近似谱图卷积模型. 自动化学报, 2021, 47(5): 1067−1076 Gong Pei-Liang and Ai Li-Hua. Two-order approximate spectral convolutional model for semi-supervised classification. Acta Automatica Sinica, 2021, 47(5): 1067−1076
[12]	张重生, 陈杰, 李岐龙, 邓斌权, 王杰, 陈承功. 深度对比学习综述. 自动化学报, 2023, 49(1): 15−39 doi: 10.16383/j.aas.c220421 Zhang Chong-Sheng, Chen Jie, Li Qi-Long, Deng Bin-Quan, Wang Jie, Chen Cheng-Gong. Deep contrastive learning: a survey. Acta Automatica Sinica, 2023, 49(1): 15−39 doi: 10.16383/j.aas.c220421
[13]	Ju W, Wang Y F, Qin Y F, Mao Z Y, Xiao Z P, Luo J Y, et al. Towards graph contrastive learning: a survey and beyond. arXiv: 2405.11868, 2024.
[14]	Zhao T X, Wang Y Q, Xu S L, Yang T C, Gao J B, Guo J P. Dual-level noise augmentation for graph clustering with triplet-wise contrastive learning. Pattern Recognition, 2025, 172: 112463 doi: 10.1016/j.patcog.2025.112463
[15]	Chen J S, Liu H P, Hopcroft J, He K. Leveraging contrastive learning for enhanced node representations in tokenized graph transformers. In: Proceedings of the 38th International Conference on Neural Information Processing Systems. Vancouver, BC, Canada: Curran Associates Inc., 2024. 85824-85845
[16]	Lee H K, Zhang Q C, Yang C, Xiong L. Node-level contrastive unlearning on graph neural networks. arXiv: 2503.02959, 2025.
[17]	Zhao Y H, Wang Y J, Wang Z K, Shan W, Huang M M, Wang X W. Graph contrastive learning with progressive augmentations. In: Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining. Toronto, ON, Canada: ACM, 2025. 2079-2088
[18]	Liang H D, Du X J, Zhu B L, Ma Z J, Chen K, Gao J B. Graph contrastive learning with implicit augmentations. Neural Networks, 2023, 163: 156−164 doi: 10.1016/j.neunet.2023.04.001
[19]	Xiao T, Zhu H S, Chen Z Y, Wang S H. Simple and asymmetric graph contrastive learning without augmentations. In: Proceedings of the 37th International Conference on Neural Information Processing Systems. New Orleans, LA, USA: Curran Associates Inc., 2023. 16129-16152
[20]	Zhao T X, Wang Y Q, Wang J L, Wang J P, Cui M L, Gao J B, et al. Hybrid-collaborative augmentation and contrastive sample adaptive-differential awareness for robust attributed graph clustering. arXiv: 2510.02731, 2025.
[21]	Wang D, Zhou W, Hu S L. Information diffusion prediction with graph neural ordinary differential equation network. In: Proceedings of the 32nd ACM International Conference on Multimedia. Melbourne, VIC, Australia: ACM, 2024. 9699-9708
[22]	Chamberlain B, Rowbottom J, Gorinova M I, Michael M. Bronstein, Webb S, Rossi E. GRAND: graph neural diffusion. In: Proceedings of the 38th International Conference on Machine Learning. Virtual Event: PMLR, 2021. 1407-1418
[23]	Matthew Thorpe, Tan Minh Nguyen, Hedi Xia, Thomas Strohmer, Andrea L. Bertozzi, Stanley J. Osher, and Bao Wang. GRAND++: graph neural diffusion with a source term. In: Proceedings of the 10th International Conference on Learning Representations. Virtual Event: OpenReview. net, 2022.
[24]	Kang Q Y, Zhao K, Ding Q X, Ji F, Li X H, Liang W F, et al. Unleashing the potential of fractional calculus in graph neural networks with FROND. In: Proceedings of the 12nd International Conference on Learning Representations. Vienna, Austria: OpenReview. net, 2024.
[25]	Zhao K, Li X H, Kang Q Y, Ji F, Ding Q X, Zhao Y N, et al. Distributed-order fractional graph operating network. In: Proceedings of the 38th Annual Conference on Neural Information Processing Systems. Vancouver, BC, Canada: Curran Associates Inc., 2024. 103442-103475
[26]	Wang J L, Guo J P, Sun Y F, Gao J B, Wang S F, Yang Y C, et al. DGNN: decoupled graph neural networks with structural consistency between attribute and graph embedding representations. IEEE Transactions on Big Data, 2025, 11(4): 1813−1827 doi: 10.1109/TBDATA.2024.3489420
[27]	Lee N, Lee J, and Park C Y. Augmentation-free self-supervised learning on graphs. In: Proceedings of the 36th AAAI Conference on Artificial Intelligence. Virtual Event: AAAI Press, 2022. 7372-7380
[28]	Zhao Y N, Ji F, Zhao K, Li X H, Kang Q Y, Liang W F, et al. Simple graph contrastive learning via fractional-order neural diffusion networks. arXiv: 2504.16748, 2025.
[29]	Kipf T N, Welling M. Semi-supervised classification with graph convolutional networks. In: Proceedings of the 5th International Conference on Learning Representations. Toulon, France: OpenReview. net, 2017.
[30]	Yu W D, Hou Z C, Liu X R. Automated polynomial filter learning for graph neural networks. In: Proceedings of the International Conference on Big Data. Washington, DC, USA: IEEE, 2024. 680-689
[31]	He M G, Wei Z W, and Wen J R. Convolutional neural networks on graphs with chebyshev approximation, revisited. In: Proceedings of the 35th Annual Conference on Neural Information Processing Systems. New Orleans, LA, USA: Curran Associates Inc., 2022. 7264-7276
[32]	Zou Z Y, Jiang Y H, Shen L, Liu J, Liu X R. LOHA: direct graph spectral contrastive learning between low-pass and high-pass views. In: Proceedings of the AAAI Conference on Artificial Intelligence. Philadelphia, PA, USA: AAAI Press, 2025. 13492-13500
[33]	Guo J P, Sun Y F, Gao J B, Hu Y L, Yin B C. Logarithmic schatten-p p norm minimization for tensorial multi-view subspace clustering. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023, 45(3): 3396−3410 doi: 10.1109/tpami.2022.3179556
[34]	Chen J Y, Lei R L, Wei Z W. Polygcl: Graph contrastive learning via learnable spectral polynomial filters. In: Proceedings of the 12nd International Conference on Learning Representations. Vienna, Austria: OpenReview. net, 2024.
[35]	Velickovic P, Fedus W, Hamilton W L, Liò P, Bengio Y, Hjelm R D. Deep graph infomax. In: Proceedings of the 7th International Conference on Learning Representations. New Orleans, LA, USA: OpenReview. net, 2019.
[36]	Zhu Y Q, Xu Y C, Yu F, Liu Q, Wu S, Wang L. Graph contrastive learning with adaptive augmentation. In: Proceedings of the Web Conference. Ljubljana, Slovenia: ACM, 2021. 2069-2080
[37]	Zhang H R, Wu Q T, Yan J C, Wipf D, Yu P S. From canonical correlation analysis to self-supervised graph neural networks. In: Proceedings of the 34th Annual Conference on Neural Information Processing Systems. Virtual Event: Curran Associates Inc., 2021. 76-89
[38]	Thakoor S, Tallec C, Azar M G, Azabou M, Dyer E L, Munos R, et al. Large-scale representation learning on graphs via bootstrapping. In: Proceedings of the 10th International Conference on Learning Representations. Virtual Event: OpenReview. net, 2022.
[39]	Wang H N, Zhang J Y, Zhu Q, Huang W, Kawaguchi K J, Xiao X K. Single-pass contrastive learning can work for both homophilic and heterophilic graph. Transactions on Machine Learning Research, 2023.
[40]	Zbontar J, Jing L, Misra I, LeCun Y, Deny S. Barlow twins: Self-supervised learning via redundancy reduction. In: Proceedings of the 38th International Conference on Machine Learning. Virtual Event: PMLR, 2021. 12310-12320
[41]	Bardes A, Ponce J, LeCun Y. Vicreg: Variance-invariance-covariance regularization for self-supervised learning. In: Proceedings of the 10th International Conference on Learning Representations. Virtual Event: OpenReview. net, 2022.