Using Output of Linear Programming-based Models

Among classification problems multi-modal problems form an important category. They naturally arise in many fields when a single class can occupy disjoint areas in the feature space. They are particularly hard to solve for traditional classification techniques. Decision tree Classifiers(DTCs) in principle are able to solve such problems by the recursive nature of their algorithm. In principle, DTCs should perform well on multi-modal data sets because they have the ability to partition each class into disjoint areas. In practice, however, DTCs can have significant difficulty with such data sets. In the presence of multimodality decision trees can became very bushy leading to high error rates. Clustering of the data before classification can lead to simpler trees and lower error rates. We apply linear programming based decision trees to multimodal problems. We present a technique that recieves input from the discrimination to alter the feature space for clustering in such a way as to improve classification accuracy.