Uncertainty Quantification in Deep Learning Context: Application to Insurance

Abstract

Nowadays, Deep learning becomes the most powerful black box predictors, which has achieved a high performance in many fields such as insurance especially in fraud detection, claims management, pricing, etc. Despite these achievements, the main interest of these classic deep learning networks is to focus only on improving the accuracy of the model without assessing the quality of the outputs. In other words, classic deep learning networks do not incorporate uncertainty information but it consists only in returning a point prediction. Knowing how much confidence there is in a prediction is essential for gaining insurers' trust in technology. In this work, we propose a solution to detect automobile insurance fraud with quantified uncertainty, our model uses two methods to quantify uncertainty. The first one is called Monte Carlo Dropout method, which is considered as an approximate Bayesian inference in deep Gaussian processes. The second is named Deep Ensembles method. These two methods mitigate the problem of representing uncertainty in deep learning without sacrificing either computational complexity or test accuracy. We found that our proposed method gives good results in comparison to the existing methods on the automobile insurance data set “carclaims.txt”.