A Novel Approach to Large-Scale Dynamically Weighted Directed Network Representation.

Area of Science:

• Big Data Analytics
• Network Science
• Machine Learning

Background:

• Dynamically Weighted Directed Networks (DWDNs) are common in big data applications.
• High-dimensional and incomplete (HDI) DWDNs arise from large-scale, dynamic interactions.
• Extracting knowledge from HDI DWDNs is challenging due to data limitations.

Purpose of the Study:

• To propose a novel Alternating Direction Method of Multipliers (ADMM)-based Nonnegative Latent-factorization of Tensors (ANLT) model.
• To efficiently handle the incompleteness and nonnegativity of HDI tensors.
• To extract hidden knowledge and behavior patterns from HDI DWDNs.

Main Methods:

• Developed a data density-oriented augmented Lagrangian function to manage HDI tensor properties.
• Implemented an ADMM framework to split optimization tasks into solvable subtasks for rapid convergence.
• Provided theoretical guarantees for the model's convergence and learning efficiency.

Main Results:

• The proposed ANLT model was tested on six real-world DWDNs.
• ANLT demonstrated significant improvements over state-of-the-art models.
• Superior performance was observed in both computational efficiency and prediction accuracy.

Conclusions:

• The ANLT model offers an effective solution for knowledge extraction from HDI DWDNs.
• The ADMM-based approach ensures efficient computation and reliable convergence.
• This research advances the analysis of complex dynamic network data in big data applications.