Research on bank fraud transaction detection based on LSTM-Focalloss

Abstract

Bank fraud transaction has brought huge losses to consumers and banks, and the original rule-based fraud detection method is not suitable for various new fraud way. According to the fact that bank fraud transaction is a typical unbalanced data classification problem, two neural network models of DNN and LSTM are established, with a new type loss function, Focalloss is used to train and test on the Kaggle's TESTIMON Dataset. As the test results, LSTM-Focalloss's network was able to detect fraud transactions significantly higher than other methods, indicating that this network model is very effective in detecting bank fraud transactions.