Using Machine Learning Methods in Behavioral Analysis for Fraud Detection on a Marketplace

In recent years, online retailers have grown in popularity, presenting their customers with a wide range of products and sellers. Along with this, the number of fraudulent activities is also growing. Attackers may have different goals, but in each case, the marketplace will suffer damage. If you look inside any such online store, it turns out that we are dealing with a large amount of data that we can use to analyze and process. By approaching the problem statement correctly and identifying the right entities and data sources, we can utilize modern machine learning techniques, including deep learning, to predict fraudulent activities. This research proposes to investigate the application of different approaches to two real-life practical tasks that exist in the largest marketplace in Russia and many other countries. Both tasks are examples of organized group fraud, both tasks are visually similar but are different in purpose and behavior pattern. The first task is the task of self-purchases, when sellers, with the help of artificial sales, try to increase their goods in recomendation to buyers. The second is the fraudulent returns task, where attackers take advantage of internal mechanisms for monetary gain, in essence ripping off the marketplace. Two different business-oriented approaches were proposed to solve both problems. Eleven algorithms were studied and the best ones were selected for use in the product solution. The obtained quality of solutions fully satisfied the business customers and allowed the marketplace to get monetary compensation in the first case and loss prevention in the second case, also in the second case the introduction of the model allowed to use a new withdrawal mechanism. And also the most valuable achievement is that the marketplace has increased the trust and safety of its users.