.png){kind=icon}
.png){kind=icon}
Return
Link prediction via robust bidirectional deep nonnegative matrix factorization
.png){kind=icon}
PRE
AI
DOI：10.1016/j.eswa.2025.128108.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Link prediction has received extensive attention due to its significant theoretical and practical value. Numerous approaches have been proposed to infer missing links or predict latent links based on observed network topology. Nonnegative Matrix Factorization (NMF) is widely employed to address the issue of link prediction because of its excellent explainability and scalability. However, most existing NMF-based link prediction methods are unidirectional shallow structural models and exhibit sensitivity towards noise. Furthermore, these methods rarely consider the nonlinear features hidden in networks. To address these challenges, we propose a novel robust bidirectional deep nonnegative matrix factorization(RBDNMF) approach for link prediction in this paper. Specifically, RBDNMF contains bidirectional deep structures which mutually guide each other to acquire more accurate node representation and captures hierarchical structural information hidden in networks. Additionally, L2,1-norm is employed to enhance the robustness and a joint kernel function is applied to explore abundant useful features, especially the nonlinear features. Finally, the convergence of objective function is strictly proved in mathematics. The effectiveness of RBDNMF is validated through a comparative analysis with twelve state-of-the-art methods across six real-world networks for link prediction task.
Keywords:
Link prediction
NMF
RBDNMF
Kernel function
L2,1-norm
Journal
IF:
7.5
Papers:
2.9W
Citations:
10.2W