Machine Learning Modeling and Evaluation for E-commerce Transaction Fraud Detection

With the continuous growth of e-commerce transaction volume, transaction fraud is becoming increasingly sophisticated in terms of concealment and harm. Addressing the characteristics of e-commerce transaction data—multi-dimensional features, complex patterns, and imbalanced categories—this paper first cleans and performs feature engineering on the raw data, constructing feature representations from dimensions such as time, transaction behavior, and account attributes. Then, under a unified process, logistic regression, gradient boosting tree, and multilayer perceptron (MLP) models are trained respectively, and a systematic comparison is conducted using 5-fold cross-validation based on six metrics: accuracy, precision, recall, F1, and ROC-AUC. Experimental results show that the gradient boosting tree outperforms the competition across all metrics (Accuracy=0.7431, Precision=0.7588, Recall=0.7120, F1=0.7343, ROC-AUC=0.8019, FLOPs= 4.16×10⁶) with smaller fluctuations between folds, indicating a stronger ability to characterize complex fraud patterns and more robust generalization performance. Logistic regression shows relatively stable performance but its overall level is limited. MLP has high precision but low recall, resulting in a limited F1 score. The research findings provide quantitative basis for the selection of fraud detection models and the optimization of risk control strategies for e-commerce platforms.