Abstract:
E-wallets and mobile money services have also been introduced that are more convenient than ever before to revolutionize the financial transactions. But more complex fraud threats are an outcome of this expansion and pose massive financial dangers to users and service provider. This has brought about the development of the necessity of designing effective and powerful fraud detecting machines. This thesis includes a comparative study in details on machine learning, ensemble, and deep learning models as far as the process of fraudulent transaction detection in e-wallet is concerned. The study uses the application of the SMM Dataset which is a simulation of mobile money behavior in terms of transactions. The methodology has a number of stages. Firstly, there is a large phase of data preprocessing, e.g., processing of missing data, coding categorical data, e.g., LabelEncoder, MinMaxScaler of numeric data. The unnecessary features are done away with in order to simplify the dataset. Lasso technique, in turn, is employed in identification of feature extraction that will forecast the most suitable predictors of fraud. This data is classified into a training set and a testing set and ratio established to be 80: 20 which was found to be the best as per the test results. The proposed models and ones tested are a variety of standard machine learning algorithms (Random Forest, K-Nearest Neighbors, Decision Tree, Logistic Regression, XGBoost), an ensemble model (Random Forest + XGBoost) and the latest models of deep learning (Artificial Neural Network, Convolutional Neural Network and Recurrent Neural Network). The standard evaluation measures (Accuracy, Precision, Recall, and F1-Score) are strictly considered in estimating the quality of every model. The results of the experiment confirm the fact that deep learning models are more effective. Convolutional Neural Network (CNN) was the most effective and highest in the accuracy of 0.9224, then came the Artificial Neural Network (ANN) with the accuracy of 0.9221. Such outcomes reveal the potential of deep learning algorithms and CNNs, in particular, in terms of revealing more complex patterns that indicate the occurrence of fraud as a possible means of enhancing the security of ewallet systems.
