Home // eKNOW 2015, The Seventh International Conference on Information, Process, and Knowledge Management // View article

LOM, a Locally Oriented Metric which Improves Accuracy in Classification Problems

Authors:
Julio Revilla
Evaristo Kahoraho

Keywords: metrics; k-NN; SVM; Riemannian; Dijkstra

Abstract:
New tools for computer automatic reasoning, casebased reasoning or data mining, require powerful artificial intelligence techniques to provide both, a high rate of precisión in their predictions, and a collection of similar past experiences that could be applied in the actual scenario. Algorithms based on Nearest Neighbors, Support Vector Machines, etc. often provide an accurate solution for classification problems, but they depend on how the similarity is measured. For this task, most of the experts employ a Euclidean metric that equally weights all attributes of the case (which is unlikely in the real world). In addition, it is well known that a correct metric choice should improve their prediction abilities and avoid the curse of dimensionality. In this paper, we present a new metric for those algorithms. It replaces the traditional Euclidean approach with a new riemannian metric that “enlarges” the space parallel to the frontier of separation between classes, thus improving classification accuracy.

Pages: 29 to 35

Copyright: Copyright (c) IARIA, 2015

Publication date: February 22, 2015

Published in: conference

ISSN: 2308-4375

ISBN: 978-1-61208-386-5

Location: Lisbon, Portugal

Dates: from February 22, 2015 to February 27, 2015