Link-Information Augmented Twin Autoencoders for Network Denoising.

Area of Science:

• Network Science
• Data Mining
• Machine Learning

Background:

• Real-world networks often contain noisy links, compromising their integrity for analysis.
• Supervised learning methods are hindered by the unreliability of observed network data due to contamination.
• Accurate network denoising is crucial for reliable data preprocessing and subsequent analysis.

Purpose of the Study:

• To develop a robust computational model for effectively removing noisy links from observed networks.
• To address the challenge of data contamination in network preprocessing.
• To accurately recover the true network structure from corrupted data.

Main Methods:

• A two-phased computational model named link-information augmented twin autoencoders is proposed.
• The model incorporates link information augmentation, link-level contrastive denoising, and link information correction.
• The approach is validated through extensive experiments on six real-world networks.

Main Results:

• The proposed model significantly outperforms comparable methods in noisy link removal.
• Experimental results demonstrate accurate recovery of the real network from corrupted observations.
• The model effectively addresses the challenges of link information augmentation, denoising, and correction.

Conclusions:

• The developed link-information augmented twin autoencoders provide a superior solution for network denoising.
• The model's effectiveness in recovering true network structures is supported by extensive empirical evidence.
• This research offers a valuable tool for preprocessing real-world network data contaminated with noisy links.