Abstract: To detect transactional fraud in financial services industry and maintain financial security, an end-to- end modeling framework, methodology, and model architecture are proposed for financial transactional data with imbalanced and discrete classes. The framework covers data preprocessing, model training, and model prediction. The performance and efficiency of different models with different numbers of features are compared and validated on a real-world dataset. The results demonstrate that the proposed approach can effectively improve the accuracy and efficiency of financial transactional fraud detection, providing a reference for financial institutions to select models with different types and numbers of features according to their own optimization goals and resource constraints. Tree-based models excel with over 200 features in resource-rich settings, while neural networks are optimal for medium-sized feature sets (100 ~200). Decision trees or logistic regression are suitable for small feature sets in resource-constrained, long-tail scenarios. 

Key words: fraud detection, artificial intelligence, model selection, machine learning, deep learning