On Reducing Misclassifications and Error on CART Models using Rtosynr Dataset for Improved Debit Card Fraud Detection
International journal of research studies in computer science and engineering · 2019
This work describes a way of reducing the number of misclassifications in Classification and Regression Tree (CART) classifiers for debit card fraud detection using the Real-to-Synthetic-Real (RtoSynR) model; thereby increasing the classification accuracy of the classifier.RtoSynR model involves the generating of synthetic transaction data from a sample of real data to augment the available real data during the training of the model.The joining of the generated synthetic data with the available real data produces the RtoSynR dataset which is used to train and improve the classification accuracy of the classifier.The RtoSynR dataset is updated first by the customer feedback and then by the modeling and generating of newly observed transaction patterns (fraudulent and non-fraudulent patterns) to solve the problem of class imbalance in the training set and concept drift in machine learning which results in underfitting and overfitting classifiers.The detailed algorithm for RtoSynR implementation was described and implemented.Result of the tests shows an improvement on the classification accuracy of the classifiers developed with RtoSynR dataset over those developed with real dataset.