A technique of dynamic feature selection using the feature group mutual information

Abstract

Knowledge discovery from raw data is very important to non-experts and even experts who feel difficulty in expressing their skills in machine interpretable forms. However, real world data often contain some redundant or unnecessary features, and if they are directly used, the quality of the seeking knowledge may be much degraded. Here, a new technique of dynamically selecting features is suggested. Contrary to the static feature selection, this scheme selects each new feature based on its correlation with the previously selected features. In addition, this scheme does not require setting any threshold, which would be too difficult to decide. Experiments have been conducted for some real world domains in terms of tree sizes and test data error rates. The results show the soundness of this scheme.