Specific-class distance measures for nominal attributes

Article type: Research Article

Authors: El Hindi, Khalil

Affiliations: Computer Science Department, College of Computer and Information Sciences, King Saud University, Riaydh, Saudi Arabia E-mails: khindi@ksu.edu.sa, kelhindi@gmail.com

Abstract: The classification accuracy of many machine learning methods depends upon their ability to accurately measure the similarity between different instances. Similarity is measured using a distance metric or measure. In this work, several novel distance measures for nominal values are proposed. These distance measures exploit the class of a training example against which a new instance is compared. The experiments, conducted using 50 benchmark data sets, indicate that the proposed functions are superior in many cases to the Value Difference Metric (VDM) that is widely used in instance based learning. Some of the proposed measures have proven to be less sensitive to missing values and noise in the training data sets and have maintained good classification accuracy in the presence of unknown and noisy attribute values. Like VDM, the proposed measures work only with labelled training data sets which makes them unsuitable for unsupervised learning methods.

Keywords: Machine learning, instance-based learning, kNN algorithm, lazy learners, distance functions, similarity metrics

DOI: 10.3233/AIC-130565

Journal: AI Communications, vol. 26, no. 3, pp. 261-279, 2013