FINANCIAL FRAUD PREDICTIONS IN E-COMMERCE USING MACHINE LEARNING AND DEEP LEARNING MODELS

Abstract

Nowadays it is often said that so many money related transactions are done using online services, day to day fraud has become a major issue for everyone using E-commerce services. Online fraudsters now focus on financial transactions, as old security measures cannot detect the more advanced ways they commit fraud. This paper examines the evolving issues related to E-commerce fraud through machine learning (ML) and deep learning (DL). Today, many shopping-related transactions are taken care of by online sites which has led to an increase in daily fraud faced by people making such purchases. Intelligent fraudsters now aim for financial transactions, as the older ways of detecting fraud cannot identify them. With ML and DL, this paper studies the developing trends in E-commerce fraud. The goal is to introduce flexible approaches to better detect financial frauds in real time. Over 20,000 transactions in E-commerce were used for the research because they appeared both imbalanced and unreliable. I made the training successful by first oversampling (SMOTE), undersampling the data and analyzing it using box plots to remove any outliers. For training and testing the models, six frameworks chosen are Random Forest, AdaBoost, CatBoost, XGBoost, Long Short-Term Memory and Gated Recurrent Unit. They were picked because they have managed to identify frauds in the past, mostly thanks to their ability to be decisive and observe the data accurately. It covers each phase of a modeling project, mainly focusing on handling dirty data, selecting the best features, selecting an appropriate model and measuring its accuracy, precision, recall and F1-score and area under the AUC-ROC curve. Even though Random Forest outperformed the other models regarding classifying, I find that the others are just as trustworthy. For this thesis study, we finish by addressing the main problems such as ensuring data balance, improving frameworks and introducing multiple ways to boost AI by making detection of fraudulent activities easier, letting AI explain itself and creating mixing models.