Martian Chronicles: Is MARS Better than Neural Networks?

A recently developed data mining technique, Multivariate Adaptive Regression Splines (MARS) has been hailed by some as a viable competitor to neural networks that does not suffer from some of the limitation of neural networks. Like neural networks, it is effective when analyzing complex structures which are commonly found in data, such as nonlineratities and interactions. However, unlike neural networks, MARS is not a "black box," but produces models that are explainable to management. This paper will introduce MARS by showing its similarity to an already well-understood statistical technique: linear regression. It will illustrate MARS by applying it to insurance fraud data and will compare its performance to that of neural networks.