SOFAR: Large-Scale Association Network Learning.

Area of Science:

• Statistics
• Machine Learning
• Data Mining

Background:

• Modern big data applications involve numerous responses and predictors, necessitating advanced analytical techniques.
• Understanding large-scale association networks through sparse latent factors is crucial for statistical efficiency and scientific discovery.
• Achieving both sparsity and orthogonality in factor analysis has been a significant challenge.

Purpose of the Study:

• To propose a novel method, sparse orthogonal factor regression (SOFAR), that accommodates both sparsity and orthogonality in analyzing large-scale data.
• To develop a robust framework for learning underlying association networks in big data.
• To provide theoretical guarantees and demonstrate practical utility across various machine learning tasks.

Main Methods:

• SOFAR utilizes sparse singular value decomposition with orthogonality-constrained optimization.
• The method employs convexity-assisted non-convex optimization techniques.
• Non-asymptotic error bounds are derived to characterize theoretical advantages.

Main Results:

• The proposed SOFAR method effectively integrates sparsity and orthogonality.
• Theoretical analysis provides non-asymptotic error bounds, demonstrating statistical guarantees.
• An efficient SOFAR algorithm with proven convergence properties is presented.

Conclusions:

• SOFAR offers a powerful approach for uncovering latent structures in large-scale datasets.
• The method demonstrates broad applicability in both unsupervised and supervised learning, including biclustering and sparse factor analysis.
• Simulations and real-world examples validate the computational and theoretical advantages of SOFAR.