1.School of Mathematics and System Science, Xinjiang University, Urumqi Xinjiang 830017, China
2.School of Biology and Engineering, Guizhou Medical University, Guiyang Guizhou 561113, China
Show less
文章历史+
Received
Accepted
Published
2026-03-31
2026-04-30
2026-07-25
Issue Date
2026-10-14
PDF (2788K)
摘要
加权符号有向图因可通过符号、方向和权重建模复杂系统相互作用,受到广泛关注.将基于拉普拉斯矩阵的谱图卷积方法应用于此类图研究极具价值,兼具简洁性与可解释性.但该应用面临三大挑战:拉普拉斯矩阵易丧失半正定性、可谱分解性和特征值有界性,限制了适用范围;模型性能对边数量和权重分布敏感,稳定性不足;计算成本较高、效率受限.为此,本文提出新型磁符号拉普拉斯矩阵,并基于该矩阵设计加权符号有向图卷积网络(Weighted Signed Directed Graph Convolutional Network, WSDGCN).理论推导表明,该矩阵适配加权符号有向图时可保留拉普拉斯矩阵的核心优良性质,实现差异化拓扑表征的同时,对边权重表现出较好的鲁棒性.在合成及真实数据集上的节点分类与链路预测实验表明,本文方法与其他方法相比具有较为优异的性能.
Abstract
Weighted signed directed graphs have attracted widespread attention for their ability to model interactions in complex systems via signed, directed and weighted edges. Applying Laplacian-based spectral graph convolution methods to such graphs is of great value, as these methods are both concise and interpretable. However, such applications face three major challenges: the Laplacian matrix tends to lose positive semi-definiteness, spectral decomposability and eigenvalue boundedness, thus limiting its scope of application; model performance is sensitive to the number of edges and weight distribution of datasets, leading to poor stability; and the computational cost is high with limited efficiency. To address these issues, this paper proposes a novel magnetic signed Laplacian matrix and designs a Weighted Signed Directed Graph Convolutional Network (WSDGCN) based on this matrix. Theoretical derivation demonstrates that this matrix preserves the core superior properties of the Laplacian matrix when adapted to weighted signed directed graphs. While achieving differentiated topological characterization, it exhibits favorable robustness to edge weights. Experiments on node classification and link prediction tasks conducted on both synthetic and real-world datasets verify that the proposed method achieves superior performance compared with other competing methods.
BackstromL, LeskovecJ .Supervised random walks:Predicting and recommending links in social networks[C/OL]//WSDM,2011:635-644.
[2]
ZouX H .A survey on application of knowledge graph[J].Journal of Physics:Conference Series,2020,1487:012016.
[3]
KashyapS, KumarS, AgarwalV,et al .Protein protein interaction network analysis of differentially expressed genes to understand involved biological processes in coronary artery disease and its different severity[J].Gene Reports,2018,12:50-60.
[4]
MaY, TangJ L .Deep learning on graphs[M].Cambridge:Cambridge University Press,2021.
[5]
WuZ H, PanS R, ChenF W,et al .A comprehensive survey on graph neural networks[J].IEEE Transactions on Neural Networks and Learning Systems,2021,32(1):4-24.
VeličkovićP, CucurullG, CasanovaA,et al .Graph attention networks[C/OL]//ICLR,2018:1-12.
[11]
HamiltonW L, YingR, LeskovecJ .Inductive representation learning on large graphs[C/OL]//NIPS,2017:1025-1035.
[12]
DefferrardM, BressonX, VandergheynstP .Convolutional neural networks on graphs with fast localized spectral filte-ring[C/OL]//NIPS,2016:3844-3852.
[13]
KipfT N, WellingM .Semi-supervised classification with graph convolutional networks[C/OL]//ICLR,2017:1-14.
[14]
KimJ, ParkH, LeeJ E,et al .SIDE:Representation learning in signed directed networks[C/OL]//WWW,2018:509-518.
[15]
DerrT, MaY, TangJ L .Signed graph convolutional networks[C/OL]//ICDM,2018:929-934.
[16]
LiY, TianY, ZhangJ W,et al .Learning signed network embedding via graph attention[C/OL]//AAAI,2020:4772-4779.
[17]
HuangJ J, ShenH W, HouL,et al .Signed graph attention networks[C/OL]//ICANN,2019:566-577.
[18]
HuangJ J, ShenH W, HouL,et al .SDGNN:Learning node representation for signed directed networks[C/OL]//AAAI,2021:196-203.
[19]
ChengH, HeC B, LiuH,et al .Community detection based on directed weighted signed graph convolutional networks[J].IEEE Transactions on Network Science and Engineering,2024,11(2):1642-1654.
[20]
KunegisJ, SchmidtS, LommatzschA,et al .Spectral analysis of signed graphs for clustering,prediction and visua-lization[C/OL]//SDM,2010:559-570.
[21]
DittrichT, MatzG .Signal processing on signed graphs:Fundamentals and potentials[J].IEEE Signal Processing Magazine,2020,37(6):86-98.
[22]
ZhangX T, HeY X, BrugnoneN,et al .MagNet:A neural network for directed graphs[C/OL]//NeurIPS,2021:1-13.
[23]
SinghR, ChenY X .Signed graph neural networks:A frequency perspective[C/OL]//NeurIPS,2023:1-19.
[24]
HeY X, PerlmutterM, ReinertG,et al .MSGNN:A spectral graph neural network based on a novel magnetic signed Laplacian[C/OL]//LoG,2022:1-39.
[25]
FioriniS, ConiglioS, CiavottaM,et al .SigMaNet:One Laplacian to rule them all[C/OL]//AAAI,2023:7568-7576.
[26]
KoT, ChoiY, KimC K .A spectral graph convolution for signed directed graphs via magnetic Laplacian[J].Neural Networks,2023,164:562-574.
[27]
FioriniS, ConiglioS, CiavottaM,et al .Graph learning in 4D:A quaternion-valued Laplacian to enhance spectral GCNs[C/OL]//AAAI,2024:12006-12015.
[28]
GolubG H, van LoanC F .Matrix computations[M].4th ed.Baltimore:Johns Hopkins University Press,2013.
[29]
KumarS, SpezzanoF, SubrahmanianV S,et al .Edge weight prediction in weighted signed networks[C/OL]//ICDM,2016:221-230.
[30]
KunegisJ, LommatzschA, BauckhageC .The Slashdot Zoo:Mining a social network with negative edges[C/OL]//WWW,2009:741-750.
[31]
GuhaR, KumarR, RaghavanP,et al .Propagation of trust and distrust[C/OL]//WWW,2004:403-412.
[32]
HeY X, ReinertG, WangS C,et al .SSSNET:Semi-supervised signed network clustering[C/OL]//SDM,2022:244-252.