Dataset
European cardholder transactions
The dataset contains 284,807 transactions made by European cardholders over two days in September 2013. There are 492 fraud cases, so fraud accounts for only 0.172% of all records.
CSV workflow
Batch score transactions
Upload a CSV with `Time`, `Amount`, and `V1` through `V28`. If the file includes `Class` or `y_true`, the batch summary reports precision and recall.
Held-out test set
Evaluation dashboard
Model comparison
Cross-validation folds
XGBoost classification report
XAI layer
Feature importance and explanations
Global SHAP importance
XGBoost gain importance
Method
Feature engineering and XGBoost
The raw PCA features are combined with time and amount transformations: hour, day, cyclic hour features, log amount, zero-amount flag, scaled time, and scaled amount.
Metric choice
PR-AUC over raw accuracy
Because the classes are highly imbalanced, the project emphasizes AUPRC, precision, recall, F1, and MCC instead of relying on plain accuracy.
Free hosting
Static Hugging Face Space
This deployed version uses the free Static Space SDK. The model runs in the browser from the exported XGBoost JSON file, so no paid Gradio CPU, paid GPU, hosted database, or external inference endpoint is required.
Hugging Face rejected free Gradio CPU creation for this account with a payment-required response, so the
app was converted to a static browser implementation to honor the no-paid-services requirement.