Fraud Detection Project

Objective

Detect fraudulent financial transactions using machine learning techniques.

Dataset

Kaggle Fraud Detection Dataset (~636,262 transactions) with features including:

• step: timestep of the transaction
• type: type of transaction (PAYMENT, TRANSFER, CASH_OUT, DEPOSIT, etc.)
• amount: transaction amount
• nameOrig: sender account
• oldbalanceOrg: initial balance of sender
• newbalanceOrig: new balance of sender
• nameDest: receiver account
• oldbalanceDest: initial balance of receiver
• newbalanceDest: new balance of receiver
• isFraud: target variable (1 if fraud, 0 otherwise)
• isFlaggedFraud: flagged by the system (not always accurate)

Tools & Libraries

• Python 3.11
• Pandas, NumPy
• Matplotlib, Seaborn
• Scikit-learn (Logistic Regression)
• Streamlit (for app deployment)

Feature Engineering

• balanceDiffOrig = oldbalanceOrg - newbalanceOrig
• balanceDiffDest = newbalanceDest - oldbalanceDest
• Focus on TRANSFER and CASH_OUT transactions

Model

• Logistic Regression with class_weight='balanced'
• Input features: numeric (amount, balances) + categorical (type)
• Preprocessing: StandardScaler for numeric, OneHotEncoder for categorical

Model Performance

• F1-score: 0.82
• Confusion matrix available in notebook
• High precision and recall achieved for fraud class

Visualization

• Transaction distribution by type
• Fraud rate by transaction type
• Boxplots and histograms for transaction amounts
• Correlation heatmap

Streamlit App

• Users input transaction details (type, amount, balances)
• Predicts if transaction is fraudulent
• Displays results with st.success or st.error

Files in Repository

• fraud_detection.ipynb → Jupyter Notebook with analysis and model training
• fraud_detection_pipeline.pkl → Trained logistic regression pipeline
• streamlit_app.py → Streamlit interface for predictions
• screenshots/ → app screenshots

Key Insights

• Only 0.13% of transactions are fraudulent → highly imbalanced dataset
• TRANSFER and CASH_OUT transactions are most prone to fraud
• Feature engineering improves model accuracy
• Logistic Regression is interpretable and effective for baseline predictions