MLP-ARD vs. Logistic Regression and C4.5 for PIP Claim Fraud Explication

In this paper we demonstrate the explicative capabilities of multilayer perceptron neural networks (MLP) with automatic relevance determination (ARD) weight regularization for personal injury protection (PIP) automobile insurance claim fraud detection. The ARD objective function hyperparameter scheme provides a means for soft input selection as it allows to determine which predictor variables are most informative to the trained MLP. The MLP (hyper-)parameters are trained using MacKay’s (1992;1994) evidence framework for classification implementation of Bayesian learning on a data set of closed PIP insurance claims from accidents that occurred in Massachusetts during 1993. We adhere to an experimental strategy of using the aggregated decision from a ten-fold cross-validated ensemble for robust predictor importance assessment. The findings of MLP-ARD are then compared to the predictor importance evaluation from popular logistic regression and (smoothed+curtailed) C4.5 decision tree classifiers trained on the same data and using an identical experimental setup. The results show good agreement between logistic regression and MLP-ARD, while somewhat less between C4.5 and MLP-ARD.