MEM: a new mixed ensemble model for identifying frauds

Abstract

In the social security system, there still exist wilful insurance frauds. In this paper, to address the insufficient stability and randomness of the traditional insurance fraud evaluation model, we propose a new classifier called mixed ensemble model (MEM). Based on the principle of ensemble learning, MEM combines several different individual learners and uses Q statistical methods to evaluate diversity. MEM has been tested on two fraud related datasets to compare with three state-of-the-art classifiers: neural network, naive Bayes and logistic regression. The experimental results show that MEM performs better than the other three classifiers in both datasets under the four measures: accuracy, recall, F-value and kappa. MEM can be a useful method for the detection of social insurance fraud.