طرق الكشف عن الاحتيال في بطاقة الائتمان - التنقيب عن البيانات

With the rapid advancement of technology, the world is turning to credit cards rather than cash in their daily lives, which opens the door to numerous new possibilities for dishonest people to use these cards in an unethical manner. Global card losses are likely to hit $35 billion by 2020, according to Nilson research. To safeguard the protection of these credit card customers, the credit card issuer should provide a service that protects consumers from any danger they may encounter (Dal Pozzolo et al., 2017).
The dataset for this study was gathered through research cooperation between Worldline and the Université Libre de Bruxelles's Machine Learning Group on the issue of big data mining and fraud detection. It is made up of numerous transactions made by European cardholders in September of 2013. After PCA transformation, the information is presented as numerical variables to ensure user confidentiality and identification. It is made up of the time between transactions and the quantities of money involved in the transactions (Anis, M., & Ali, M., 2017).
The credit card dataset is substantially unbalanced since it contains more legal transactions than fraudulent ones. That is, without identifying a fraudulent transaction, the prediction will have a very high accuracy score. Class distribution, i.e., sampling minority classes, is a preferable technique to deal with this type of situation. In minority sampling, class training examples can be increased in proportion to the majority class to boost the algorithm's chances of the right prediction (Brownlee, J., 2020).
Several studies are being conducted to identify fraudulent transactions using deep neural networks. These models, on the other hand, are computationally costly and perform better on bigger datasets. This strategy may provide excellent results, as seen by certain studies, but we can obtain the same, or even better, results with fewer resources. So, our major objective is to demonstrate that with proper preprocessing, several machine learning algorithms may provide satisfactory results(Kazemi, Z., & Zarrabi, H.,2017).
As a result, the (Ada Boost) algorithm, according to our findings, brings the highest results, i.e., better determines whether transactions are fraudulent or not. This was assessed using a variety of criteria, including recall, accuracy, and precision. For this type of circumstance, having a high recall value is crucial. The significance of feature selection and dataset balance in producing significant results has been proven.