Fraud Detection System

This repository contains a fraud-detection workflow built around the IEEE-CIS / Vesta e-commerce fraud dataset. It includes a model-training notebook, a Streamlit app for live prediction, the raw CSV inputs, and the saved model artifacts needed to run inference.

Project Overview

The training pipeline uses a hybrid machine-learning approach:

• Missing values are filled with -999.
• Categorical features are label-encoded.
• Features are scaled with StandardScaler.
• KMeans is used as an additional engineered feature.
• Class imbalance is handled with SMOTE.
• Multiple models are trained and compared, including Logistic Regression, Decision Tree, Random Forest, Optuna-tuned XGBoost, and LightGBM.

The notebook evaluates models with F1 score, ROC-AUC, confusion matrix, and classification report, then saves the inference artifacts used by the app.

Files In This Folder

• app.py - Streamlit web app for live fraud prediction.
• Fraud-Detection.ipynb - End-to-end training and evaluation notebook.
• transaction.csv - Main transaction dataset.
• identity.csv - Identity metadata joined with the transaction data.
• xgb_model.pkl - Saved XGBoost model used by the app.
• lgbm_model.pkl - Saved LightGBM model used by the app.
• scaler.pkl - Saved scaler used during preprocessing.
• kmeans.pkl - Saved KMeans clustering model used as a hybrid feature.

What The App Does

The Streamlit app in app.py provides:

• An introduction page describing the fraud-detection problem, dataset, and pipeline.
• A live prediction demo where users enter transaction details and select either the XGBoost or LightGBM model.
• A probability-based risk view with fraud probability, confidence, and a visual risk meter.

Requirements

Install the project dependencies from requirements.txt:

pip install -r requirements.txt

If you want to work with the notebook, also make sure Jupyter is installed:

pip install notebook ipykernel

How To Run

1. Set up a virtual environment

python -m venv .venv
.venv\\Scripts\\activate

2. Install dependencies

pip install -r requirements.txt

3. Launch the Streamlit app

streamlit run app.py

The app expects xgb_model.pkl, lgbm_model.pkl, scaler.pkl, and kmeans.pkl to be in the same folder as app.py.

4. Open the notebook

Run Fraud-Detection.ipynb in Jupyter or VS Code to inspect the full training pipeline, reproduce the model comparisons, and regenerate the saved artifacts if needed.

Data Notes

• The dataset is highly imbalanced, with fraud representing a small fraction of transactions.
• transaction.csv is large, so some editors may not preview it fully.
• The app is designed for interactive demonstration, not for production risk scoring without further validation and monitoring.

Troubleshooting

• If the app shows missing model-file errors, confirm the .pkl files are in the project root.
• If package import errors appear, verify the virtual environment is activated and all dependencies are installed.
• If the notebook fails on memory-heavy operations, reduce the dataset size for local experimentation or run it in an environment with more RAM.