On the Impact of Distance Metrics in Instance-Based Learning Algorithms

• Autor(a) principal: Lopes, Noel
• Data de Publicação: 2015
• Outros Autores: Ribeiro, Bernardete
• Tipo de documento: Artigo
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: http://hdl.handle.net/10314/3247; https://doi.org/10.1007/978-3-319-19390-8
• topic: Distance metrics; Instance-based learning; Incremental learning; Nearest Neighbor; Incremental Hypersphere Classifier (IHC)

In this paper we analyze the impact of distinct distance metrics in instance-based learning algorithms. In particular, we look at the well-known 1-Nearest Neighbor (NN) algorithm and the Incremental Hypersphere Classifier (IHC) algorithm, which proved to be efficient in large-scale recognition problems and online learning. We provide a detailed empirical evaluation on fifteen datasets with several sizes and dimensionality. We then statistically show that the Euclidean and Manhattan metrics significantly yield good results in a wide range of problems. However, grid-search like methods are often desirable to determine the best matching metric depending on the problem and algorithm.