A Machine-Learning Framework: Integrating MLP And SMOTE For Deceitful Payment Classifications

In this study, Multiple Layer Perceptron’s (MLP) integrating Synthetic Minority Oversampling Technique (SMOTE) was offer as potential solution to the despicable problem of payment fraud. This model was design and realized on Google Colab platform embedding GPU, where Tensorflow was fit for deep learning (DL), and Scikit learn for machine learning (ML) model respectively. Python was the modeling language. Firstly, baseline experiment was conduct on orthodox ML models of Random Forest (RF), Logistics Regression (LR), Isolation Forest (iforest) with MLP for feature selection and engineering. While, probing kaggle dataset. This dataset is highly imbalance, a bait to the ML challenges suffering in the baseline trial. However, resampling method of SMOTE was applied in the second trial to balance the dataset, address over-fitting with under-fitting problem and improved on the utilized model performances. In testing for these models’ efficacy, confusion matrix and performance evaluation metric were explored. This revealed the outcome of the balancing model trial, where the proposed MLP+SMOTE exercises superclass performance against other models. Presenting accuracy score of 0.95%, Error Rate of 0.05%. Recall of 0.98%, least False Positive Rate of 0.07%, True Negative Rate of 0.93%, Precision of 0.93%, Prevalence of 0.46%, Null Error Rate of 100%, Cohen kappa of 0.42%, F1-Score of 0.95% and Matthews Correlation Coefficient of 0.91% respectively. This result validates the developed model as amazing in performance, when compared with benchmark studies; and it is promising in the classification of deceitful payment transaction.