MMFS-CF: A personalized data-driven credit card fraud detection model based on multi-modal multi-objective feature subset selection

Credit card fraud detection (CCFD) is a critical research direction in the field of financial risk prevention and control, aiming to protect the interests of consumers and financial institutions by identifying suspicious transactions. Nevertheless, due to the high privacy and sensitivity of some transaction features, as well as the difficulty in extracting some transaction features or the high cost of obtaining them, the requirement for personalized transaction features in the CCFD scenario cannot be met. Additionally, the original transaction data often has irrelevant and redundant features, which are not conducive to the improvement of CCFD performance. Therefore, we present a personalized data-driven CCFD model named MMFS-CF based on multi-modal multi-objective feature subset selection, which focuses on two key optimization objectives: minimizing the count of transaction features and maximizing CCFD performance (i.e., AUC). Specifically, we develop a dynamically adaptive guidance vector mechanism through the construction of a multi-subpopulation collaborative evolution framework. This mechanism adaptively directs subpopulations to converge toward unexplored regions within the decision space, guided by real-time population density information. Furthermore, it integrates two key components: a genetic operation enhancement strategy by embedding guidance vectors to improve population diversity and accelerate convergence, and a guidance vector-driven environmental selection update mechanism aimed at refining solution quality. A key innovation lies in its personalized feature selection paradigm, enabling decision-makers to flexibly select from alternative feature subsets tailored to real-world constraints (e.g., privacy concerns or computational limitations), all of which do not affect CCFD performance. We showcase the detection capabilities of MMFS-CF using a large-scale private commerce dataset as well as four publicly available datasets. The experimental findings validate that MMFS-CF can deliver superior CCFD performance and highlight its multi-modal efficacy.