Unsupervised Prototype Reduction for Data Exploration and An Application to Air Traffic Management Initiatives

We discuss a new approach to unsupervised learning and data exploration that involves summarizing a large data set using a small set of ``representative'' elements. These representatives may be presented to a user in order to provide intuition regarding the distribution of observations. Alternatively, these representatives can be used as cases for more detailed analysis. We call the problem of selecting the representatives the Unsupervised Prototype Reduction problem. We discuss the KC-UPR method for this problem and compare it to other existing methods that may be applied to this problem. We propose a new type of distance measure that allows for more interpretable presentation of results from the KC-UPR method. We demonstrate how solutions from the Unsupervised Prototype Reduction problem may be used to provide decision support for the planning of air traffic management initiatives, and we produce computational results that compare the effectiveness of several methods in this application. We also provide an example of how the KC-UPR method can be used for data exploration, using data from air traffic management initiatives at Newark Liberty International Airport