Joint learning adaptive metric and optimal classification hyperplane.

Area of Science:

• Machine Learning
• Computer Vision
• Pattern Recognition

Background:

• Traditional metric learning often relies on k-nearest neighbors (kNN), where the choice of 'k' impacts classification generalization.
• Existing methods face challenges in jointly learning adaptive metrics and optimal classification boundaries.

Purpose of the Study:

• To propose an end-to-end metric learning framework for enhanced classification performance.
• To develop novel linear (LMML) and nonlinear (RLMML) metric learning models.
• To address optimization challenges in non-convex metric learning models.

Main Methods:

• Introduced Linear Metric Learning (LMML) to jointly learn adaptive metrics and classification hyperplanes by maximizing margin.
• Developed a nonlinear extension (RLMML) using a bound nonlinear kernel function.
• Employed half-quadratic optimization algorithms for iterative optimization of metrics and hyperplanes.

Main Results:

• The proposed LMML and RLMML models demonstrated effectiveness on various datasets.
• Numerical experiments confirmed the superior performance of the developed algorithms.
• Theoretical convergence proofs were established for the optimization algorithms.

Conclusions:

• The novel metric learning framework, including LMML and RLMML, offers an effective approach for classification tasks.
• Half-quadratic optimization provides a viable solution for optimizing complex, non-convex metric learning problems.
• The proposed models show significant feasibility and effectiveness validated by experimental results and statistical tests.