DEEP LEARNING BASED HYBRID APPROACH OF DETECTING FRAUDULENT TRANSACTIONS

As daily transactions made with credit cards have been increasing, fraudulent transactions have also continuously increased. Therefore, the importance of detecting anomalous transactions has kept rising. The given dataset, from Kaggle, consists of imbalanced data, 99.83% of normal data and 0.17% of fraud data. Therefore, in order to solve this imbalance problem, we decided to construct a fraud detecting algorithm. Through constructing a new model with a hybrid approach of deep learning and machine learning, which is composed of a Bi-LSTM-Autoencoder and Isolation Forest, we successfully detected fraudulent transactions in the given dataset. This proposed model yielded an 87% detection rate of fraudulent transactions. Compared to other models (Isolation Forest, Local Outlier, and LSTM-Autoencoder), which show 79%, 3% and 82% detection rates, respectively, our proposed model attained the highest rate. On the contrary, when evaluated by accuracy score, our proposed model did not show a higher score. Even though our model has a similar accuracy score compared to other models and does not implement  the Variational Autoencoder for feature selection, this model could potentially be utilized as an effective process to detect fraudulent transactions, especially with the number of global cases increasing along with the need for productivity, quicker detection.