Survey and experimental study on metric learning methods.

Area of Science:

• Machine Learning
• Data Science

Background:

• Distance metric learning enhances methods like k-nearest neighbors (kNN) by learning data-dependent metrics.
• Existing surveys offer limited theoretical analysis or narrow experimental comparisons.
• A comprehensive review and experimental evaluation of metric learning methods are needed.

Purpose of the Study:

• To review influential distance metric learning methods from 2003-2017.
• To propose a taxonomy for categorizing these methods.
• To experimentally compare selected methods for performance evaluation.

Main Methods:

• Literature review of classical and influential distance metric learning algorithms.
• Taxonomic classification into five categories: pairwise cost, probabilistic framework, boost-like, advantageous variants, and specific applications.
• Comprehensive experimental study comparing selected methods on accuracy, distance change, and kNN neighbor size.

Main Results:

• Categorization of methods based on distinct characteristics.
• Experimental comparison revealing the performance of different metric learning approaches.
• Analysis of the relationship between distance modification, accuracy, and kNN parameters.

Conclusions:

• Metric learning is crucial for improving distance-based algorithms.
• The proposed taxonomy provides a structured overview of existing methods.
• Experimental results offer insights into method selection for specific applications.