A Convex Model for Support Vector Distance Metric Learning.

Area of Science:

• Machine Learning
• Pattern Recognition
• Data Mining

Background:

• Existing k-Nearest Neighbors Distance Metric Learning (k NN DML) methods require storing all training data and are sensitive to the 'k' parameter.
• Current Support Vector Machine (SVM) based DML methods lack convexity, hindering global optimization.
• A convex approach for SVM-based DML has not been explicitly proposed.

Purpose of the Study:

• To propose a novel convex model for Support Vector Distance Metric Learning (CSV-DML).
• To replace the computationally intensive k NN model with a more efficient SVM model in DML.
• To develop a DML method that utilizes kernel functions and ensures global convergence.

Main Methods:

• Developed Convex Support Vector Distance Metric Learning (CSV-DML) using kernel transformation for nonlinear mapping.
• Learned a Mahalanobis distance metric on kernel-transformed instances.
• Employed generalized block coordinate descent for global optimization.
• Introduced a parameter reduction scheme for high-dimensional kernel spaces.

Main Results:

• CSV-DML effectively replaces the k NN model with an SVM classifier.
• The proposed method converges to a global optimum due to its convex formulation.
• Experimental results demonstrate that CSV-DML outperforms previous DML methods.
• The parameter reduction scheme effectively manages feature dimensions.

Conclusions:

• CSV-DML provides a robust and efficient convex alternative for distance metric learning.
• The method leverages kernel functions for enhanced performance.
• CSV-DML offers significant improvements over existing distance metric learning techniques.