Constrained K-Means Clustering Validation Study

Abstract

Machine Learning (ML) is a growing topic within Computer Science, and has applications in many fields. One classical problem in ML is the question of separating data, and this process is now known as clustering. One interesting application of data clustering is making insurance adjustments for hail damage to crops. Our project is a validation study of, “Constrained K-means Clustering with Background Knowledge by Wagstaff et. al. Here we show that a modified k-means clustering approach can outperform more general unsupervised learning when some domain information about the problem is available. In Wagstaff et.al, machine learning was applied to the problem of predicting what soybean production would have been if hail had not damaged crops. Because this is an estimation that many crop insurance agents have had to make, it is a relatively large data set that has been labeled by human experts. Our data suggests that k-means clustering augmented with domain information can be a time efficient means for segmenting data sets. Our validation study focused on six classic data sets and does not consider the GPS data of the original study. We have published our code onto a SWOSU Github repository to enable other researchers to use our code as a starting point. Validation studies such as this provide great learning opportunities for students interested in working with Machine Learning, Artificial Intelligence, and other related applications.