Robust Graph Learning From Noisy Data.

Area of Science:

• Machine Learning
• Data Science
• Computer Vision

Background:

• Automated graph learning shows promise for clustering and semisupervised tasks.
• Real-world data often contains noise, leading to unreliable learned graphs.
• Existing methods struggle with corrupted data, impacting downstream task performance.

Purpose of the Study:

• To develop a novel robust graph learning scheme for reliable graph construction from noisy data.
• To improve the performance of clustering, semisupervised classification, and data recovery tasks.
• To address the limitations of current graph learning techniques in handling real-world data corruption.

Main Methods:

• Propose a robust graph learning scheme that adaptively removes noise and errors from raw data.
• View the model as a robust manifold regularized robust principle component analysis (RPCA).
• Leverage graph smoothness assumptions for enhanced low-rank recovery and RPCA for improved graph construction.

Main Results:

• The proposed model significantly boosts performance in data clustering, semisupervised classification, and data recovery.
• Enhanced low-rank recovery and improved graph construction contribute to overall performance gains.
• Experiments demonstrate superior performance over state-of-the-art methods on various tasks.

Conclusions:

• The novel robust graph learning scheme effectively learns reliable graphs from noisy data.
• The method offers significant improvements in clustering, semisupervised classification, and data recovery.
• The approach provides a robust alternative to existing graph learning techniques for real-world applications.