CSTG: An Effective Framework for Cost-sensitive Sparse Online Learning.

Area of Science:

• Machine Learning
• Data Mining
• Optimization

Background:

• Sparse online learning and cost-sensitive learning are established research areas.
• Limited research exists on the combined study of these two fields.
• High-dimensional data streams with skewed distributions present significant challenges.

Purpose of the Study:

• To introduce a novel framework for cost-sensitive sparse online learning.
• To address the challenges of high-dimensional data streams with skewed distributions.
• To simultaneously balance misclassification cost and sparsity.

Main Methods:

• Developed a framework extending the Truncated Gradient (TG) method.
• Formulated a new convex optimization problem.
• Theoretically analyzed regret and cost bounds.

Main Results:

• The proposed algorithm effectively balances misclassification cost and sparsity.
• Theoretical analysis demonstrates merit compared to related approaches.
• Large-scale empirical comparisons on eight real-world datasets show encouraging performance against five baseline methods.

Conclusions:

• The developed cost-sensitive sparse online learning framework is effective for high-dimensional, skewed data streams.
• The method offers theoretical advantages in balancing competing objectives.
• Empirical results validate the practical performance of the proposed approach.