Learning With Selected Features.

Area of Science:

• Machine Learning
• Statistical Learning
• Big Data Analytics

Background:

• The proliferation of big data presents challenges for traditional machine learning algorithms.
• Classical kernel-based regularized least-squares (RLS) faces computational and storage limitations, hindering its application in big data scenarios.

Purpose of the Study:

• To introduce a scalable algorithm, Learning with Selected Features (LSF), designed to overcome the computational bottlenecks of RLS.
• To derive optimal learning rates and conditions for kernel and center selection to ensure algorithmic optimality.

Main Methods:

• The study proposes a subsampling-based approach for the LSF algorithm.
• Theoretical analysis is conducted to determine optimal learning rates and selection criteria.
• Numerical experiments, including simulations and real-world data, are used for verification.

Main Results:

• LSF effectively reduces the computational burden associated with RLS.
• The proposed method demonstrates near-optimal learning rates and provides conditions for guaranteed optimality.
• Experimental results validate the theoretical assertions across various datasets.

Conclusions:

• The LSF algorithm offers a viable solution for applying RLS in the era of big data.
• LSF achieves significant computational savings while maintaining high generalization ability.
• This work facilitates the use of powerful RLS methods on large-scale datasets.