Generalized Low-Rank Update: Model Parameter Bounds for Low-Rank Training Data Modifications.

Area of Science:

• Machine Learning
• Computational Statistics

Background:

• Incremental machine learning (ML) is crucial for dynamic datasets.
• Existing low-rank update methods are limited to linear estimators.
• Updating ML models efficiently after data changes is computationally challenging.

Purpose of the Study:

• To develop a generalized low-rank update (GLRU) method for ML.
• To extend efficient model updating beyond linear estimators.
• To enable efficient cross-validation and feature selection.

Main Methods:

• Developed the generalized low-rank update (GLRU) method.
• Formulated ML methods as regularized empirical risk minimization.
• Extended low-rank update framework to support vector machines and logistic regression.

Main Results:

• The GLRU method efficiently updates ML models with incremental data changes.
• Achieved computational complexity proportional to the number of data changes.
• Demonstrated efficiency in cross-validation and feature selection tasks.

Conclusions:

• GLRU significantly enhances the efficiency of incremental ML.
• The method is applicable to a broader range of ML algorithms.
• GLRU offers a practical solution for model selection in dynamic environments.