Robust Subspace-Constrained Quadratic Models for Low-Dimensional Structure Learning
A new robust subspace-constrained quadratic model (SCQM) is proposed for learning low-dimensional structure from high-dimensional data. Building on the SQMF framework, it handles various noise distributions including generalized Gaussian and radial Laplace, improving robustness under heavy-tailed and light-tailed noise. A gradient-based algorithm with backtracking line-search ensures stable convergence. Sensitivity analysis of ℓp^p and ℓ2 loss functions reveals their behavior under different noise. Numerical experiments validate the approach.
Key facts
- Proposes robust subspace-constrained quadratic model (SCQM)
- Builds upon subspace-constrained quadratic matrix factorization (SQMF)
- Accommodates generalized Gaussian and radial Laplace noise
- Enhances robustness under heavy-tailed and light-tailed noise
- Develops gradient-based algorithm with backtracking line-search
- Presents sensitivity analysis of ℓp^p and ℓ2 loss functions
- Extensive numerical experiments corroborate the method
- Published on arXiv with ID 2605.20300
Entities
Institutions
- arXiv