Explainable Fraud Detection AI System in Financial Sector

Abstract

Digital payment services now handle millions of transactions each day, where even a tiny fraction of fraud causes major financial losses and undermines customer trust. This paper investigates how to accurately detect fraudulent transactions in a Kaggle financial payment services dataset and to understand which transaction behaviors make payments appear risky. The original dataset is pre-processed through outlier removal, label encoding, standardisation and two stages balancing strategy combining random down sampling of the majority class with SMOTE oversampling. Nine numeric and categorical features are then used as inputs to classifiers including LR, KNN, DT, RF, SVM, GNB, AdaBoost, Bagging, Voting and Stacking under four settings. They are no dimensionality reduction (NoDR), UMAP, NCA and PLS-DA respectively. Hyperparameters are tuned with GridSearchCV using both 70/30 train–test split and 10-fold cross-validation. Model performance is evaluated with accuracy, precision, recall, F1-score, specificity, ROC-AUC, PR-AUC and training time. SHAP is applied to interpret feature importance and local decisions. Without dimensionality reduction Bagging, Random Forest and Stacking achieve 99.3 to 99.4% accuracy and F1, with ROC-AUC and PR-AUC close to 99.95% on the test set and similarly strong cross-validation scores. UMAP and NCA preserve high performance in lower dimensional spaces. While PLS-DA gives moderate but consistent results. The study shows that ensemble models on the original feature space provide highly accurate and stable fraud detection on this dataset. SHAP analysis reveals that source and destination balances, transaction amount and type are the most influential features.