Detecting fraudulent transactions presents two primary obstacles, the presence of imbalanced datasets and the sheer volume of feature categories. The scarcity of fraudulent instances within the dataset often results in these cases being erroneously classified as noise, consequently introducing bias towards non-fraudulent cases in the detection outcomes. Simultaneously, the extensive array of feature categories poses a formidable challenge in identifying crucial variables, thereby limiting the efficacy of detection algorithms. This research contributes to the field of fraud transaction detection by proposing a multifaceted approach that addresses the challenges of imbalanced data and feature complexity. Through down-sampling, dimension reduction, model evaluation, and precision enhancement, the study aims to improve the accuracy and efficiency of fraud detection models, with the goal of reducing false alarms and financial losses in real-world scenarios.