Атты III халықаралық ғылыми-тәжірибелік конференция жинағЫ
"SCIENCE AND EDUCATION IN THE MODERN WORLD
жүктеу/скачать 8,11 Mb. Pdf көрінісі
|
| "SCIENCE AND EDUCATION IN THE MODERN WORLD:
CHALLENGES OF THE XXI CENTURY" NUR-SULTAN, KAZAKHSTAN, JULY 2019 37 processing data streams will be required. The ability to transfer the model and the integration of decision-making logic into operating systems is of paramount importance for preventing fraud on the widest scale - given the scope of fraudulent transactions. It is very important to be able to explain what exactly the machine learning system does. The system or solution, the principles of which are clear to the user, is usually called "white boxes". As a rule, methods and models of machine learning are impenetrable "black boxes" - the user does not know exactly how they work. It is very difficult (almost impossible) to explain to analysts why they got one or another result or solution. There are many approaches to adding evaluation cards, taking into account local linear approximation, as well as creating textual parts and graphic visualizations. All these are only approximate values, but they give users an idea of the machine learning model, as well as useful recommendations on the study of fraudulent activities. Everything is changing, and we need to be able to adapt to change. Continuous monitoring of machine learning-based fraud detection systems is an indisputable key to success. As the models and the underlying data change, the quality of the input data deteriorates and the overall system performance decreases. This problem is not only peculiar to machine learning systems, but also to rule-based systems. However, new methods of machine learning are able to effectively adapt to new, as yet unknown, patterns. This makes it possible to reduce the number of necessary measures (although not to exclude them all) in retraining and evaluating the operation of the machine learning system. An effective monitoring system actively explores the data that enters the system, evaluates the forecasts and explanations generated by the machine learning model, and also notifies administrators of changes in data and statistics trends before radical changes affect the performance of the entire company. For one of the financial institutions, combating fraudulent transactions was an intractable problem. It was necessary not only to identify illegal operations, but also to provide a high level of customer service. An effective fraud detection system should not block legitimate customer transactions. The financial institution sought to modernize the existing rule-based system and achieve an optimal balance between control functions and customer service. For this, his representatives turned to SAS. Their goal was to use the capabilities of neural networks to create two separate systems for evaluating fraudulent actions: A system for calculating the likelihood that a customer's account is under the control of fraudsters. A system for calculating the likelihood that a single transaction is fraudulent. Thanks to this approach, the financial institution was able to identify transactions in the amount of almost $ 1 million per month, which were erroneously identified as fraudulent, as well as identify transactions in the amount of $ 1.5 million per month, which were fraudulent, but were not detected by the previous system. Our solution not only helped the company more effectively detect fraud, it also made it possible to significantly increase customer satisfaction by easing tensions in their relationship with the company. How? Thanks to a significant improvement in transaction confirmation procedures and increased fraud detection. Successful machine learning programs always involve varying degrees of continuous experimentation. It is not enough just to create a machine learning model and send it ―to float freely‖. Scammers are smart, and technology is constantly changing. The presence of an isolated environment, a sandbox, in which data scientists can experiment with various methods, data and technologies to combat fraud, in this case becomes a critical condition for the implementation of the most important programs. Investments that are aimed at optimizing the work and increasing the productivity of data scientists involved in the detection and prevention of fraud will pay for themselves almost instantly. |