Quadratic Matrix Factorization With Applications to Manifold Learning.

Area of Science:

• Machine Learning
• Data Science
• Computer Vision

Background:

• Matrix factorization is a common technique for low-rank data modeling.
• Existing methods often struggle with the inherent curvature of real-world datasets.
• Manifold learning aims to uncover the underlying structure of high-dimensional data.

Purpose of the Study:

• To introduce Quadratic Matrix Factorization (QMF) for learning curved manifolds.
• To provide an optimization algorithm and convergence analysis for QMF.
• To adapt QMF for robust manifold learning applications.

Main Methods:

• Developed a Quadratic Matrix Factorization (QMF) framework.
• Proposed an alternating minimization algorithm for QMF optimization.
• Introduced a regularized QMF to prevent overfitting and discussed parameter tuning.

Main Results:

• QMF effectively captures the curved structure of data manifolds.
• The proposed alternating minimization algorithm converges theoretically.
• Regularized QMF demonstrates superior performance on synthetic and real-world datasets.

Conclusions:

• QMF offers a powerful alternative to linear methods for manifold learning.
• The proposed regularization technique enhances model robustness.
• Experiments confirm QMF's effectiveness on diverse datasets like MNIST and cryo-EM.