L1 -norm low-rank matrix factorization by variational Bayesian method.

Area of Science:

• Computer Vision
• Pattern Recognition
• Machine Learning

Background:

• L1-norm low-rank matrix factorization (LRMF) is crucial for computer vision and pattern recognition.
• Existing LRMF methods struggle with noise and outliers, impacting generalization.
• A robust and efficient LRMF approach is needed.

Purpose of the Study:

• To develop a novel hierarchical Bayesian generative model for L1-norm LRMF.
• To design an efficient mean-field variational inference method for parameter estimation.
• To enhance the robustness and generalization capability of LRMF.

Main Methods:

• Constructed a hierarchical Bayesian generative model for L1-norm LRMF.
• Employed mean-field variational inference with closed-form solutions.
• Adaptive weighting of matrix elements to mitigate noise and outliers.
• Incorporated L2-regularization for parameter generalization.

Main Results:

• The proposed method adaptively weights matrix elements, suppressing noise and outliers.
• Variational Bayesian inference provides statistically guaranteed parameter regularization.
• Demonstrated superior robustness and efficiency on synthetic and real data.
• Achieved better prediction accuracy on unobserved data than state-of-the-art methods.

Conclusions:

• The Bayesian LRMF approach offers enhanced robustness and generalization.
• Adaptive weighting effectively handles noisy and outlier-prone data.
• The method provides a statistically sound and efficient solution for L1-norm LRMF problems.