Gaussian Process for Regression in Business Intelligence: A Fraud Detection Application

Abstract

Business Intelligence (BI) systems are designed to provide information to support the decision making process in companies and governmental institutions. In this scenario, future events depend on the decisions and on the previous events. Therefore, the mathematical analysis of past data can be an important tool for the decision making process and to detect anomalies. Depending on the amount and the type of data to be analyzed, techniques from statistics, Machine Learning (ML), data mining and signal processing can be used to automate all or part of the system. In this paper, we propose to incorporate Gaussian Process for Regression (GPR) in BI systems in order to predict the data. As presented in this work, fraud detection is one important application of BI systems. We show that such application is possible with the use of GPR in the predictive stage, considering that GPR natively returns a full statistical description of the estimated variable, which can be used as a trigger measure to classify trusted and untrusted data. We validate our proposal with real world BI data provided by the Brazilian Federal Patrimony Department (SPU), regarding the monthly collection of federal taxes. In order to take into account the multidimensional structure of this specific data, we propose a pre-processing stage for reshaping the original time series into a bidimensional structure. The resulting algorithm, with GPR at its core, outperforms classical predictive schemes such as Artificial Neural Network (ANN).