Predictor Variables' Influence on Classification Outcome in Insurance Fraud Detection

Abstract

In fraud detection paradigms, the role of predictor variables cannot be overemphasized particularly when analytical tools – such as statistical, machine learning and artificial intelligent tools are employed. These variables or attributes are used to organize records of data in database tables. The combination of the values of these attributes usually affects which class of a target variable a record or an observation would belong. In this paper, we propose an algorithm and employ spreadsheet 'count' 'count if' and 'filtering' functionalities (techniques) to take toll on how the individual attribute may affect the prediction of the class of an observation in an insurance dataset of 5000 observations. The analysis showed that indeed, the individual predictor attribute affects the outcome of the target variable (legal or fraudulent) differently.