Boxes Are Better Than Circles For Classification

摘要

The design of a classifier usually has the important step of attribute or feature selection. A computationally tractable scheme almost always relies on a subset of attributes that optimize a certain criterion is chosen. The result is usually a 'good' sub-optimal solution. We have previously shown that it is possible to define the space for each class by finding the space where the classes overlap. Using related methods, we now show that it is possible to directly define the area for each cluster in terms of an n-dimensional box. As with our previous algorithm, this classifier is transparent, and the approach compares favorably with previous approaches both in accuracy and efficiency. It has the added advantage of being able to classify when one class in entirely within another class, such as a box inside a box.