Machine Learning in the Problem of Financial Fraud Detection

The final qualifying work is devoted to the development of approaches necessary for detecting compromised payment transactions when making purchases in the classifieds and paying for them with a bank card. The approaches are based on the search for signs of fraud using machine learning and other analytical models. As part of the research work, an analysis of the subject area was carried out, namely, ad aggregators on the Russian market were considered, the features of detecting fraud with bank transactions were derived, the main principles of classifieds and ways of monetization were considered, from which three main fraud scenarios were derived. In addition, an analysis of banking transactions and the main types of payments on the Internet was carried out. The scheme of operation of a single-stage payment, a two-stage payment, and the operation of the 3-D Secure protocol are considered. When forming requirements based on fraud scenarios, hypotheses were formulated for the data, as well as requirements for the resulting model. In addition, six basic classification algorithms were described: Neural Networks, Decision Trees, Logistic Regression, K Nearest Neighbors, Support Vector Machine, and Naive Bayes. Their advantages and disadvantages are analyzed. The model development tools were selected, which include Python 3.0 and Jupyter Notepad. They were later applied to the received data. The developed approaches were tested on the basis of real classifieds data containing depersonalized information about transactions: buyer, ad, transaction data. According to the results of testing, these models demonstrated a sufficiently high predictive ability, at which the model was able to detect approximately 92% of fraudulent transactions and reach an F-measure level of more than 0.99. The resulting complex approaches are intended to be used as transaction verification algorithms for systems of the Fraud Prevention and Fraud Detection classes.