Kullback-Leibler Divergence Metric Learning.

Area of Science:

• Machine Learning
• Information Theory
• Data Science

Background:

• Kullback-Leibler divergence (KLD) is crucial for measuring distribution similarity.
• Metric learning aims to find optimal distance metrics for datasets.
• Existing KLD methods may lack expressivity for complex data distributions.

Purpose of the Study:

• To develop an enhanced KLD metric learning method.
• To improve the expressivity of data distribution representations.
• To achieve superior performance in classification tasks.

Main Methods:

• Extended conventional KLD with a linear mapping for optimization.
• Formulated KLD metric learning as a positive-definite matrix manifold minimization problem.
• Developed an intrinsic steepest descent method to preserve metric structure during optimization.

Main Results:

• The proposed method effectively enhances data distribution expressivity.
• Achieved superior performance compared to ten popular metric learning approaches.
• Demonstrated effectiveness on 3-D object and document classification tasks.

Conclusions:

• The novel KLD metric learning approach offers significant improvements.
• The method provides a robust way to learn optimal KLD-type metrics.
• This work advances metric learning for classification and related applications.