Using Clustering for Supervised Feature Selection to Detect Relevant Features

摘要

In many applications in machine learning, large quantities of features and information are available, but these can be of low quality. A novel filter method for feature selection for classification termed COLD is presented that uses class-wise clustering to reduce the dimensionality of the data. The idea behind this approach is that if a relevant feature would be removed from the set of features, the separation of clusters belonging to different classes will deteriorate. Four artificial examples and two real-world data sets are presented on which COLD is compared with several popular filter methods. For the artificial examples, only COLD is capable to consistently rank the features according to their contribution to the separation of the classes. For the real-world Dermatology and Arrhythmia dataset, COLD demonstrates the ability to remove a large number of features and improve the classification accuracy or, at a minimum, not degrade the performance considerably.