Abstract:
Today’s Tax fraud embraces various new means to commit fraud including declaring wrong information, underpaying tax due and carrying out financial businesses without considering legal frameworks. Like any other tax, Value-Added-Tax (VAT) is vulnerable to fraud which affects the growth of any country due to its numerous advantages and benefits. Recognising noncompliance for VAT’s taxpayers is a weighty as well as challenging matter for Rwanda
Revenue Authority (RRA), since there is a huge volume of VAT returns received daily and
monthly that need complex techniques in order to discover new insights and analyse it
effectively. Hence the need for a valuable intelligent tool to fight against fraud known as data mining to extract for patterns in massive volume of VAT data and automatically distinguish fraudulent patterns from legal ones. The main purpose of this present study is to analyse relationship between VAT’s patterns, build and evaluate a data mining model for fraud detection on VAT historical data for RRA.The proposed solution used SQL queries to analyse patterns according to RRA business rules, and the model architecture is designed to reason using the classification techniques Naïve Bayes, Kneighbors and Decision Tree to classify the status for taxpayer VAT’s compliance with two categories that are fraudulent and legitimate. Furthermore,
the model performance is presented and compared. The classification results generated by our model on each technique are compared with respect to the performance measures such as accuracy, precision, recall, F1-Score and ROC curves. Generally, both algorithms showed a significant accuracy but the best performing being Naïve Bayes with 98% of accuracy. The developed data mining model is promising to effectively detect VAT fraud and therefore help to generate knowledge that can be used in the audit work performed by the RRA for feature decision making.
