Aprendizaje supervisado de funciones de distancia: estado del arte

Abstract

The selection of a suitable distance function is fundamental to the instance-based learning algorithms. Such distance function influences the success or failure of these algorithms. Recently it has been shown that even a simple linear transformation of the input attributes can lead to significant improvements in classification algorithms as k-Nearest Neighbour (k-NN). One of the main applications of these algorithms is in the hybridization with instance-based learning algorithms and in that sense learning a distance metric for the application at hand and not using a general distance function; which has been shown to improve the learning results. This article presents an overview of distance metric learning, and it is modeled as an optimization problem. It then discusses different approaches to learning from the availability of information in the form of restrictions, focusing on supervised approach, and under it the global and local ones. Further models and strategies of the most representative algorithms of each approach are described.

Keywords

• classification
• distance metric learning
• k-Nearest Neighbours