Application of appropriate similarity measures into machine learning

Abstract

Learning using similarity is a common strategy for both human learning and machine learning. Psychologists and statisticians have proposed various kinds of similarity measures for both kinds of learning such as probability and distance representation. However, how similarity should be used for learning appropriately remains a controversial topic. According to some researchers such as Heit and Rubinstein, one should not make use of a single fixed similarity measure for different kind of property as concern of human learning. In this paper, we aim to support this proposal in machine learning framework and explain why different similarity measures should be used when accounting for different properties by finding suitable and reasonable similarity measures for a specific human voice recognition task. Based on the experiment results, it is concluded that appropriate similarity measures should be chosen for different kinds of properties according to the real application environment and requirement.