Lebichot, BertrandPlancq, HadrienHadrienPlancq2025-05-142025-05-142025-05-142022https://hdl.handle.net/2078.2/27064Payment card fraud is a significant problem for business owners, card issuers, and transactional services firms, resulting in significant financial losses each year. With the growing volume of data created by payment card transactions, it has become impossible for a human analyst to identify fraudulent patterns in transaction datasets, which are frequently characterized by a huge number of samples and multiple dimensions. As a result, during the last decade, the development of payment card fraud detection systems has shifted to approaches based on machine learning techniques, which automate the process of detecting fraudulent patterns from vast amounts of data. However, not all steps can be automated. Some conclusions cannot yet be made efficiently by machine learning. More specifically, ML algorithms have a lot of difficulty dealing with class imbalance. Indeed, transaction data contains far more genuine transactions than fraudulent ones: in a real-world dataset, the percentage of fraudulent transactions is often considerably under 1%. Learning from imbalanced data is difficult since most learning algorithms struggle with big class disparities. Class imbalance necessitates the adoption of additional learning procedures such as sampling or weighting. In this work, we will try to improve the efficiency of existing machine learning models as well as to create new ones with the goal of pushing for the total automation of this process, removing the need for human operators. Moreover, this study aims to try to draw conclusions on the avenues to be explored in future research concerning the detection of fraud in transactions.Unbalanced classificationFraud detectiontext::thesis::master thesisthesis:35295