Providing and evaluating a comprehensive model for detecting fraudulent electronic payment card transactions with a two-level filter based on flow processing in big data
Hamid Banirostam, Touraj Banirostam, Mir Mohsen Pedram, Amir Masoud Rahmani
Abstract
Previous research on fraud detection modeling is often based on a single algorithm, optimizing categories and clusters to find fraudulent patterns that they have provided unsupervised or supervised methods alone and within the framework of Hadoop. The proposed model, a model based on big data analysis extracts important features of user behavior patterns such as time, device type, values, and type of transaction, and their behavioral modeling. By creating different profiles for users, threshold values will be set for each of them. The proposed model for real-time fraud detection of electronic cash payment cards includes two fast and explicit filters. Fast filtering is the combination of the Hidden Markov Model of the first order and Self Organizing Maps (SOM) for fast transaction processing and the Baum-Welch algorithm to find the local Maximum Likelihood. Explicit filter model training is a combination of multilayer Perceptron Neural Network algorithms and Logistic Regression to create a cardholder profile and measure the amount of new transaction deviation from the created profile with the reduction mapping approach or parallel execution of the model. Also, transactions performed for aggregation filters and prediction of the output results of the Schaffer Demister function with the parameter θ in order to detect the degree of deviation of the transaction behavior from the normal state according to the profile of the customer and the maximum difference between two consecutive sequences of observations is used. Finally, the fraudulent or non-fraudulent label on the transaction will be applied and added to the relevant database for storage for later use. According to the model simulation results, the proposed Accuracy, Precision, Recall, and F1-Score criteria are 0.999, 0.9834, 0.7906, and 0.9214, respectively. Simulation results show that the proposed model will perform better in each of the compared criteria against each of the single algorithms.