Evaluation of classifiers for feature selection problems

The goal of instance-based classification is to generate a "good" classifier from a labelled set of examples. The classifier can then be used to classify unlabelled instances by predicting their class based on the features presented to the classifier. We present an empirical method of evaluating the learning task involved in object classification based on subsets of features. Also, we present comparisons between three classifiers, decision trees, Euclidean decision tables and K-means. The purpose of the experiments is to show how a classifier can learn from the data and how much of the data is necessary to achieve good levels of learning. Preliminary results show that the accuracy of classifiers is not improved dramatically with the use of more than one half of the number of features.