Sci Rep. 2025 Aug 15;15(1):29959. doi: 10.1038/s41598-025-15783-2.
ABSTRACT
With the rapid development of the global financial market, the problem of financial fraud is becoming more and more serious, which brings huge economic losses to the market, consumers and investors and threatens the stability of the financial system. Traditional fraud detection methods based on rules and statistical analysis are difficult to deal with increasingly complex and evolving fraud methods, and there are problems such as poor adaptability and high false alarm rate. Therefore, this paper proposes a financial fraud detection model based on Stacking ensemble learning algorithm, which integrates many basic learners such as logical regression (LR), decision tree (DT), random forest (RF), Gradient Boosting Tree (GBT), support vector machine (SVM) and neural network (NN), and introduces feature importance weighting and dynamic weight adjustment mechanism to improve the model performance. The experiment is based on more than 1 million real financial transaction data. The results show that the Stacking model is significantly superior to the traditional single model in accuracy (95%), recall (93%) and F1 score (94%), and has stronger generalization ability and stability. Although the Stacking model has challenges in computing cost and delay, its advantages in fraud detection accuracy and robustness make it a powerful tool for financial institutions to improve their risk control ability. In the future, its real-time adaptability can be further optimized through online learning and incremental update mechanism.
PMID:40817346 | DOI:10.1038/s41598-025-15783-2
