Artificial Intelligence for Detecting Irregularities in Banking Operations

Abstract

Fraudulent banking operations bring several problems related to financial safety, reliability of operations, and risk management. The goal of this research is to detect fraudulent operations in banks applying artificial intelligence and machine learning techniques. The methodology includes data preprocessing of banking transactions, missing values' imputation, encoding of categorical features, normalization of numerical features, and random under sampling due to class imbalance problem. Transaction features such as transaction amount, account balance, transaction type, geographical location, transaction frequency, and history of fraud detection are used to discover possible frauds. Machine learning classifiers are used to differentiate between normal operations and fraudulent transactions. The innovation consists in combining transaction-related behavioral characteristics, class imbalance handling, and classifier comparison in fraud recognition framework. There are 50,000 samples in the dataset; among them, there are 47,573 normal transactions, 2,423 fraud transactions, and four samples having no fraud label. According to the given confusion matrix, the framework demonstrates 98.33% accuracy, 98.66% precision, 98.00% recall, and 98.33% F1-score. It proves good classification results for fraud recognition tasks. To get actual values of ROC and AUC, predictions must be calculated for various classification thresholds.