期刊名称：International Journal of Computers and Communications
印刷版ISSN：2074-1294
出版年度：2012
卷号：6
期号：2
页码：128-139
出版社：University Press
摘要：Instance-Based learners are simple, yet, effective learners. They classify a new instance based on the k most similar instances which makes them sensitive to noise in training data sets. Obtaining good classification accuracy may, therefore, require cleaning the data sets using labor-extensive or computationally expensive data cleaning procedures. In this work, we present some Bayesian-based instance weighting techniques to make such learners more tolerant to noise. The basic idea is that typical or classical instances should be given more weight or voting power than less typical or noisy instances. We present three techniques to determine instance weights that are based on the conditional probability of an instance belonging to its actual class and not to another class. Our empirical results using the kNN algorithm shows that all presented techniques are effective in making the kNN more tolerant to noise. These results suggest that these techniques can be used with instance based learners instead of more expensive data cleaning procedures.
