Performance analysis of stacking neural network and machine learning model for detecting fraudulent transaction

Abstract

Transaction fraud has become a fast growing issue in the world of modern technology which has become a serious threat to the financial sectors. Although these fraudulent actions have several categories or type but online financial fraud has been a dominant issue so far. In reality, a profoundly precise procedure of identification of fraudulent transaction is required since it is causing a extensive wealth related depletion. Therefore, we have conducted research on financial fraud record using machine learning models and proposed a procedure for precise misrepresentation recognition dependent on the points of interest and restrictions of each exploration. In our initial stage, we implemented machine learning classifiers such as Logistic Regression, K-Nearest Neighbor, Support Vector Classifier, Na¨ıve Bayes, Gaussian Na¨ıve Bayes Classifier, Random Forest Classifier, Extra Tree Classifier, Neural Network and Adaptive Boosting to see how all of them performs separately. We also balanced the dataset that we used in order to overcome the overfit issue. Then again we tested the above mentioned classifiers on the balanced dataset. After that we tried our final step which is the implementation of Stacking technique. The accuracy that stacking method came up with were the best along with very less overfitting issues since K-fold cross validation was applied. To further boost the accuracy, we implemented Grid Search Hyperparameter tuning to get the best possible outcome at a much lower error rate. Therefore, to give a superior outcome for different sorts of online money transaction frauds, we have been keen on working with this issue and build a solid and defensive platform for safe transactions of money.