Attribute Selection for Classification

Abstract

The selection of attributes used to construct a classification model is crucial in machine learning, in particular with instance similarity methods. We present a new algorithm to select and rank attributes based on weighing features according to their ability to help class prediction. The algorithm uses the same structure that holds training records for classification. Attribute values and their classes are projected into a one-dimensional space, to account for various degrees of the relationship between them. With the user deciding on the degree of this relation, any of several potential solutions can be used as criterion to determine attribute relevance. This low complexity algorithm increases classification predictive accuracy and also helps to reduce the feature dimension problem.