A comparison of unsupervised learning techniques for detection of medical abuse in automobile claims.

摘要

Automobile claims abuse is a widespread problem that costs insurer and consumers billions of dollars per year in lost profits and higher premiums. Due to logistical and legal complications, however, many insurers are reluctant to classify formally abuse and fraud. Unfortunately, this removes the ability to perform supervised learning since the true classification of abuse is not known. Insurers are thus forced to employ unsupervised learning techniques to detect abusive claims.;The purpose of this project is to compare the effectiveness of three unsupervised learning methods on automobile claims medical abuse in one anonymous U.S. state. The analysis is performed on a collection of abusive behavioral patterns recommended by seasoned claims adjustors. Of the three unsupervised learning methods, two of these---K-Means and hierarchical clustering-are commonly used in multivariate statistics. The third method, PRIDIT (principal component analysis of relative to an identified distribution), is a novel technique that has the potential of not only accurately classifying abuse, but also categorizing the importance of each pattern. The merits and drawbacks of all three techniques are analyzed in this paper.