Robust Matrix Completion With Column Outliers.

Area of Science:

• Data Science
• Machine Learning
• Linear Algebra

Background:

• Matrix completion aims to reconstruct low-rank matrices from partial data.
• Existing methods struggle with corrupted samples (outliers) and dynamic, streaming data.
• Batch-based methods are inefficient for out-of-sample or vector completion problems.

Purpose of the Study:

• Develop a novel robust matrix completion model.
• Address challenges of corrupted data and the need for online/vector completion.
• Enable efficient handling of streaming data and outlier detection.

Main Methods:

• Proposed a robust matrix completion model with an efficient optimization technique.
• Utilized singular value decomposition (SVD) once per iteration for a thin matrix.
• Developed a vector completion model for out-of-sample prediction using online matrix completion principles.

Main Results:

• The proposed method demonstrates superior performance on streaming data with column outliers compared to traditional methods.
• The model effectively detects outliers within incomplete datasets.
• Validated through extensive experiments on both synthetic and real-world data.

Conclusions:

• The novel robust matrix completion model effectively handles corrupted data and streaming environments.
• The developed vector completion model efficiently addresses out-of-sample prediction.
• The method offers a robust solution for real-world applications requiring outlier detection and dynamic data handling.