Analyzing a Method to Determine the Utility of Adding a Classification System to a Sequence for Improved Accuracy
Date of Award
Master of Science in Applied Mathematics
Department of Mathematics and Statistics
Christina M. Schubert Kabban, PhD
Frequently, ensembles of classification systems are combined into a sequence in order to better enhance the accuracy in classifying objects of interest. However, there is a point in which adding an additional system to a sequence no longer enhances the system as either the increase in operational costs exceeds the benefit of improvements in classification or the addition of the system does not increase accuracy at all. This research will examine a utility measure to determine the valid or invalid nature of adding a classification system to a sequence of such systems based on the ratio of the change in accuracy to the increase in operational costs. Three general classification sequence strategies defined on a two-class population outcome will be examined: Believe the Positive, Believe the Negative and Believe the Extreme. Through simulation, this research will identify which characteristics of the individual classification systems and the sequence have the greatest impact on the utility measure and provide guidance on the threshold value for the utility measure that differentiates between when the addition of a system to the sequence may be useful (valid) and when it is not (invalid). This work expands upon known accuracy and cost equations for each of the different sequential strategies in order to generalize them to any fixed number of classification systems in a sequence. From these accuracy and cost calculations, the utility measure can be computed for different scenarios and recommendations are made as to the characteristics that enhance the utility of adding additional systems to a sequence in order to improve classification accuracy.
DTIC Accession Number
Pamilagas, Kevin S., "Analyzing a Method to Determine the Utility of Adding a Classification System to a Sequence for Improved Accuracy" (2019). Theses and Dissertations. 2193.