Fraud Shield — Credit Card Fraud Detection App (6 ML models compared)

Show the Community!
, , ,
1

Hey everyone! :waving_hand:

live_prediction
live_prediction1555×783 86.7 KB

I just deployed Fraud Shield, an end-to-end credit card fraud detection app — sharing it here in case it’s useful to anyone working on classification problems or looking for Streamlit UI ideas.

:link: Live app: https://musfirah-credit-card-fraud-detection.streamlit.app/

The problem

The dataset (Kaggle’s classic credit card fraud dataset) has 284,807 transactions, but only 492 (0.17%) are fraud. That imbalance makes accuracy a useless metric — a model that never flags fraud would still score 99.8% accuracy while catching nothing.

What I built

  • Trained and tuned 6 classification models with GridSearchCV: Logistic Regression, SGD, Decision Tree, Random Forest, HistGradientBoosting, and Naive Bayes
  • Handled the imbalance with class_weight='balanced' instead of oversampling
  • Evaluated on F1, ROC-AUC, Precision, and Recall instead of accuracy
  • Random Forest won (F1: 0.775, ROC-AUC: 0.983, Recall: 82.7%)
  • Extracted real feature importance from the winning model so the UI’s interactive sliders reflect what the model actually learned, not generic guesses

The Streamlit app has 5 pages:

  • :house: Home — best model snapshot + scores
  • :magnifying_glass_tilted_left: Single Prediction — check a transaction manually, or load a real one from the dataset to see guaranteed-correct model behavior
  • :open_file_folder: Batch Prediction — upload a CSV, get predictions for every row, download results
  • :bar_chart: Model Performance — full comparison across all 6 models with charts
  • :bust_in_silhouette: About Me — a bit about who built it

One Streamlit feature I leaned on a lot: st.session_state to let users toggle between “load a real example” mode and manual slider mode without losing state on rerun. Happy to share more detail on that if anyone’s running into similar session_state gotchas.

Which model do you think works best for imbalanced data? Let me know in the replies!

Would love feedback, especially on the UI/UX side — first real Streamlit app I’ve deployed! :folded_hands:

:star: **Support the Project:**
If you find this project or the Streamlit UI implementation useful, please consider giving it a **:star:** on the GitHub repository! It keeps me motivated to build more open-source tools.

#MachineLearning #FraudDetection streamlit #Python #ScikitLearn