A perturbation-based framework for link prediction via non-negative matrix factorization.

Area of Science:

• Network Science
• Data Mining
• Machine Learning

Background:

• Link prediction aims to infer missing connections in networks.
• Existing methods struggle with noise and irregularities in real-world networks.
• Non-negative Matrix Factorization (NMF) is a powerful tool for dimensionality reduction and feature extraction.

Purpose of the Study:

• To propose a robust perturbation-based framework for missing link prediction.
• To address the limitations of existing methods in handling network noise and irregular links.
• To enhance the accuracy and reliability of link prediction in complex networks.

Main Methods:

• A perturbation-based framework utilizing Non-negative Matrix Factorization (NMF).
• Automatic determination of the optimal number of latent features (inner rank) using the Colibri method.
• Repeatedly perturbing the training set to generate multiple network instances.
• Obtaining common basis and coefficient matrices via NMF to create a similarity matrix for prediction.

Main Results:

• The proposed framework demonstrates competitive performance across fifteen real-world networks.
• Experimental results show superior or comparable accuracy compared to state-of-the-art link prediction methods.
• Identified correlations between method performance and network statistics, highlighting consistency in high-precision methods.

Conclusions:

• The perturbation-based NMF framework effectively predicts missing links in noisy networks.
• The method offers a robust alternative to existing link prediction techniques.
• Network characteristics influence the performance of different link prediction strategies.