Roberto W. Menezes,  Alexandre Plastino,  Bianca Zadrozny,  Rafael B. Pereira,  Luiz Henrique de C. Merschmann,  Alex A. Freitas. 

Attribute selection is a data preprocessing step used to identify at­tributes relevant to the classification task. Recently, a lazy technique which postpones the choice of attributes to the moment an instance is submitted to classification was proposed. It is believed that knowledge about the attribute values of an instance may contribute to identify the best attributes. This lazy ap­proach used a measure based on the entropy concept to evaluate the quality of the attributes. In this work, we propose a measure based on the chi­square test statistic. Experimental results show that the use of lazy selection with the pro­posed measure allows to obtain a better predictive accuracy than the equivalent traditional attribute selection strategy for a significant number of the datasets.

http://www.lbd.dcc.ufmg.br:8080/colecoes/waamd/2009/008.pdf

Processando consulta.