Use of artificial intelligence algorithms to enhance fraud detection in the Banking Industry

Abstract

Fraud is among the most menacing problems with which every human society grapples, given the devastating impact on the effects. This practice refers to the deliberate use of false information to swindle another individual or organization money or property (Association of Certified Fraud Examiners, 2021). The banking industry has for decades used rule-based systems to flag fraud and human review of transactions. Rule-based systems encompass utilizing algorithms which perform a variety of detection actions that are written manually by fraud experts (Oniyilo, 2016). These systems require the manual adjustment of scenarios, which make it challenging to implicitly detect the transactional correlations that would point to fraud. Due to the inherent weaknesses of the rule-based fraud detection approach at banks and limited data that affects commonly used supervised machine learning algorithms, there is an urgent need for new detection techniques or systems that can handle the rapidly increasing fraud and money laundering incidences that adversely affect the Kenyan banking system. This research aimed to analyse and evaluate various machine learning algorithms to determine their performance in fraud detection for mobile banking transactions within the banking system. The study's objectives were to identify the data attributes that are best suited for mobile banking fraud detection machine learning algorithms and compare the performance of machine learning algorithms in fraud detection in mobile banking transactions. This study used the CRISP_DM methodology to determine the most accurate fraud detection algorithm. It was published to standardise the data mining processes over the industries. It has since evolved to be the most used methodology in mining data, performing analytics, and projects in data science. Crisp-DM follows the following general steps Business understanding, Data understanding, Data preparation, Modelling, Evaluation and Deployment. The research results demonstrated that logistic regression did not perform and indicated by the scores and did not predict any fraudulent transactions for the original unbalanced data. As such, it is therefore not recommended for fraud detection. Naïve Bayes performance based on the confusion matrix performed poorly as the algorithm predicted 89,997 false positives, 0 False negatives and 55 True negatives. While predicting the fraudulent transactions accurately, many non-fraudulent transactions were predicted to be fraudulent. These results go hand in hand with results from the scores, which demonstrated that Naïve Bayes had 0.08% accuracy. K Nearest Neighbour had the best results; the algorithm's accuracy was 99%, with 19 true negative predictions, two false positives and 8 False negatives. Therefore, KNN was identified as the preferred algorithm for fraud detection. Additionally, it is noted that when the transactional data is removed KNN, performed marginally better than when the static and scoring data from the fraud detection system are removed.