Kernelized Sparse Bayesian Matrix Factorization

Area of Science:

• Machine Learning
• Computer Vision
• Signal Processing

Background:

• Matrix factorization is key for extracting low-rank and sparse structures.
• Kernelized Matrix Factorization (KMF) incorporates side information but often requires manual tuning and faces computational challenges.
• Existing KMF models struggle with automatic rank determination and efficient computation.

Purpose of the Study:

• To develop a hierarchical Kernelized Sparse Bayesian Matrix Factorization (KSBMF) model.
• To automatically infer parameters and latent variables, including the reduced rank.
• To integrate side information effectively for improved matrix approximation.

Main Methods:

• Developed a hierarchical KSBMF model using variational Bayesian inference.
• Achieved low-rankness via sparse Bayesian learning.
• Enforced columnwise sparsity on latent factor matrices.
• Integrated KSBMF with a nonlocal image processing framework for denoising and inpainting algorithms.

Main Results:

• The KSBMF model automatically infers parameters and latent variables, including the reduced rank.
• Simultaneous achievement of low-rankness and columnwise sparsity.
• Developed novel algorithms for image denoising and inpainting based on KSBMF.
• Demonstrated superior performance of KSBMF over state-of-the-art methods in image restoration tasks.

Conclusions:

• The proposed KSBMF model offers an effective approach for incorporating side information in matrix factorization.
• KSBMF provides automatic parameter inference and rank determination, overcoming limitations of existing KMF models.
• The developed image restoration algorithms show significant improvements in denoising and inpainting under various corruption levels.