User Segmentation Based on Behavioral Analysis

In today's business landscape, the rise of e-commerce has significantly diminished transaction costs in retail, resulting in heightened competition among suppliers to capture the attention of customers and drive down prices. Consequently, there has been an emergence of customers who engage in activities that violate service terms and attempt to exploit various promotional mechanics. To counteract this behavior, various antifraud detection measures are being implemented globally. This paper delves into one potential approach for fraud prevention. The primary objective is to accurately pinpoint potentially fraudulent users without negatively impacting the majority of users utilizing the service. This paper explores strategies that have been successful in addressing similar challenges in e-commerce, banking, telecommunications, and software development fields, and proposes a combined approach tailored to the specific characteristics of the service in question. The issue of binary classification of users concerning class imbalance and feature sparsity is thoroughly analyzed. All technical criteria and performance metrics outlined as objectives have been successfully met. The thesis includes a comparative analysis of several experiments conducted with varying hyperparameters of gradient boosting classification model and sizes of accumulated markup. Furthermore, a comprehensive evaluation of the potential profitability resulting from the implementation is provided, emphasizing the necessity for conducting precise measurements through AB tests on the advertising platform. Lastly, the work concludes by suggesting further enhancements to the current project and identifying potential research avenues in related fields that could benefit from the insights gained through this project.