Fraud-Detection

🛡️ Fraud Detection ML System

An industry-grade machine learning pipeline for detecting fraudulent financial transactions in real-time. Built to mirror production-level fraud detection systems used by banks and e-commerce platforms, this project addresses severe class imbalance, engineers domain-specific fraud indicators, and compares multiple classifiers using industry-standard evaluation metrics.

---

📊 Project Demo

Dashboard	ROC & PR Curves

---

🎯 Problem Statement

Financial institutions and e-commerce platforms face increasing fraudulent transactions resulting in major financial losses. Rule-based systems fail to adapt to evolving fraud patterns. This project builds a machine learning model that identifies suspicious transactions based on behavioral signals and historical patterns.

---

🏗️ Pipeline Architecture

Raw Transaction Data
        │
        ▼
┌─────────────────────┐
│  Data Generation /  │  ← 50,000 transactions, 3.1% fraud rate
│     Ingestion       │
└────────┬────────────┘
         │
         ▼
┌─────────────────────┐
│   Preprocessing     │  ← Label encoding, StandardScaler, stratified split
└────────┬────────────┘
         │
         ▼
┌─────────────────────┐
│  Class Imbalance    │  ← SMOTE-style oversampling + class_weight='balanced'
│    Handling         │
└────────┬────────────┘
         │
         ▼
┌─────────────────────┐
│ Feature Engineering │  ← 6 domain fraud indicators
└────────┬────────────┘
         │
         ▼
┌─────────────────────┐
│   Model Training    │  ← 4 classifiers trained & compared
└────────┬────────────┘
         │
         ▼
┌─────────────────────┐
│    Evaluation &     │  ← ROC-AUC, PR-AUC, F1, Precision, Recall
│  Visualisation      │
└────────┬────────────┘
         │
         ▼
   Output Reports + Charts

---

⚙️ Features

🔬 Feature Engineering

Feature	Description
IsNightTxn	Flags transactions between 10 PM – 4 AM (high-risk window)
AmountZScore	Standardised transaction amount (detects outlier spending)
HighAmountFlag	1 if transaction amount > 95th percentile
TransactionCount_24h	Rolling count of transactions in a 24-hour window
FrequencyFlag	1 if more than 8 transactions in 24 hours
RiskScore	Composite weighted risk indicator (0–1 scale)

🤖 Models Trained

Model	Notes
Logistic Regression	Baseline linear classifier
Decision Tree	Interpretable rule-based splits
Random Forest	Ensemble — strong feature importance
Gradient Boosting	Boosted ensemble — typically best performer

📉 Class Imbalance Strategy

• SMOTE-style oversampling — synthetic minority samples generated by interpolating between existing fraud records (no external library needed)
• class_weight='balanced' — applied to all scikit-learn estimators to penalise misclassification of the minority class

---

📈 Results

Model	ROC-AUC	PR-AUC	F1	Precision	Recall
Logistic Regression	1.0000	1.0000	0.9984	1.0000	0.9968
Decision Tree	1.0000	1.0000	1.0000	1.0000	1.0000
Random Forest	1.0000	1.0000	1.0000	1.0000	1.0000
Gradient Boosting	1.0000	1.0000	1.0000	1.0000	1.0000

Note: Near-perfect scores reflect the synthetic dataset's clean separation. On real-world data (e.g. Kaggle Credit Card Fraud dataset), expect ROC-AUC in the 0.95–0.99 range with meaningful differentiation between models.

---

🚀 Getting Started

Prerequisites

pip install scikit-learn pandas numpy matplotlib seaborn

Run

https://github.com/RussanaMary/Fraud-Detection
cd fraud-detection-ml
python Fraud.py

Output

Three files are saved to the project directory:

fraud_detection_dashboard.png   ← 5-panel visual dashboard
roc_pr_curves.png               ← ROC & Precision-Recall curves
fraud_detection_report.md       ← Full project documentation

---

📁 Project Structure

fraud-detection-ml/
│
├── Fraud.py                          # Main pipeline (single file, fully self-contained)
├── fraud_detection_dashboard.png     # Generated dashboard
├── roc_pr_curves.png                 # Generated ROC/PR curves
├── fraud_detection_report.md         # Generated report
└── README.md                         # This file

---

🔄 Using Real Data

To swap in the Kaggle Credit Card Fraud dataset:

1. Download creditcard.csv and place it in the project folder
2. Replace the generate_transaction_data() call in main() with:

df = pd.read_csv("creditcard.csv")
df = df.rename(columns={"Time": "Hour"})   # align column names

The rest of the pipeline is plug-and-play.

---

🧠 Key Concepts Demonstrated

• Imbalanced dataset handling — real-world fraud datasets are typically 0.1–5% positive class
• SMOTE oversampling — synthetic sample generation without external dependencies
• Precision vs Recall tradeoff — critical in fraud: false negatives (missed fraud) vs false positives (blocking legit customers)
• Threshold tuning — default 0.5 threshold can be lowered to 0.3 to prioritise recall at the cost of precision
• Feature importance — Random Forest exposes which signals drive fraud predictions
• Multiple evaluation metrics — accuracy alone is misleading on imbalanced data; F1, PR-AUC, and ROC-AUC tell the real story

---