Article Preview
Top1. Introduction
The mobile payment system provides a platform in which smartphone users can transfer money to each other. This is performing by using codes sent through SMS. One user knows the smartphone number of the receiving user for transactions. There are various benefits of this service for users like the simplicity of transferring money. However, anomalous transactions may occur as they become very common nowadays. Financial fraud in terms of anomalous transactions can cause a loss of billions of dollars annually. It is a challenging task to detect anomalous transaction due to a large number of transaction and a variety of money laundering tricks. Therefore, there is a need to detect anomalous transactions to prevent financial fraud. According to (Lopez-Rojas, 2016) the ever changing techniques used by{Lopez-Rojas, 2016 #46} money launderers make this problem interesting to study. The mobile money transfer is providing anonymity, speed and portability as compared to the traditional banking system raising risks of non-banking elements which requires new approached to detect fraudulent transactions (Novikova & Kotenko, 2019).
A variety of machine learning techniques are being used to detect anomalous transactions in different domains (Gray & Debreceny, 2014). Mobile money surveillance systems (Martin, 2019) can be utilized in different ways to detect money laundering and fraudulent transactions. A variety of other techniques are also proposed by (Fahmi, Hamdy, & Nagati, 2016) and a number of essential countermeasures against money laundering were suggested by (Lopez-Rojas & Axelsson, 2012).{Lopez-Rojas, 2012 #45}{Lopez-Rojas, 2012 #45} {Lopez-Rojas, 2012 #45} The technique of data mining is (Phua, Lee, Smith, & Gayler, 2010) able to detect anomalous transactions in the mobile payment system. It has three general stages that are used to detect such transactions Discovery, Predictive and Forensic model. The discovery model is the procedure of appearing in a record to discover unseen models with no pre-determined design or assumption regarding what the models might be. This model is the procedure of captivating these examples to exposed to the records and then use these records to forecast the future. The forensic model is used to process and then apply these patterns to extract inconsistent or abnormal data elements. Bayesian network(Phua et al., 2010) is used to detect such transaction in the mobile payment system. An excellent formation and constraint algorithm are essential to finding out the model. In the experiments that we are conducting, we used the optimal re-insertion algorithm to analyze the formation and after that do an utmost probability opinion of the network consideration. When we able to build the model and then to experiment any of the instances from the database we discover its whole instances likewise specified the probability model. The tests of instances that have especially minimum likewise they are further standard as variances. Statistical approaches (Bolton & Hand, 2002) are also used to detect anomaly detections from various fields such that credit card, network intrusion, Money laundering, Telecommunication, etc. Fraud detection can be supervised or unsupervised. A supervised method is used to construct a model that yields a suspicion score for new cases.
The lack of original data set is the main issue but we have generated synthetic data set by using the PaySim simulator. The aim of this thesis develops a technique to detect anomalous transactions in the mobile payment system. We performed experiments on the financial transactional dataset using eight different data mining classification algorithms: Decision table, Ripper, Decision Stump, Partial decision tree, J48, Naïve Bayes, Locally weighted learning and K-Nearest Neighbor. Performance evaluation of classification algorithms is measured by using evaluation metrics: accuracy, precision, F-score, recall, and specificity. We also performed a comparison of applied classification algorithms. The developed technique identifies anomalous transactions and non-anomalous transactions in the financial dataset. The comparative result shows that J48 (C4.5) classifier achieved high accuracy and precision rate.