Distance metric learning based on the class center and nearest neighbor relationship

Area of Science:

• Machine Learning
• Data Mining
• Pattern Recognition

Background:

• Existing distance metric learning methods often rely solely on class centers or nearest neighbor relationships.
• Overlapping class centers pose challenges for traditional metric learning approaches.
• Effective distance metrics are crucial for accurate classification and data analysis.

Purpose of the Study:

• To propose a new distance metric learning method (DMLCN) that integrates class center and nearest neighbor information.
• To address challenges posed by overlapping class centers by introducing clustering.
• To develop a robust classification decision rule and an iterative optimization algorithm.

Main Methods:

• The proposed Distance Metric Learning based on Class center and Nearest neighbor (DMLCN) method clusters overlapping classes.
• It learns a distance metric that enforces proximity to cluster centers and maintains nearest neighbor relationships within receptive fields.
• Multiple metrics (MMLCN) are introduced for complex data, learning a local metric per cluster center.

Main Results:

• The DMLCN and MMLCN methods achieve simultaneous intra-class compactness and inter-class dispersion.
• An iterative optimization algorithm is developed with theoretical analysis of convergence and complexity.
• Experiments demonstrate the feasibility and effectiveness of the proposed methods on diverse datasets.

Conclusions:

• The proposed DMLCN and MMLCN methods offer a significant advancement in distance metric learning.
• These methods effectively characterize local data structures, leading to improved classification performance.
• The developed techniques provide a robust framework for handling complex and overlapping data distributions.