Project 3: Machine Learning System for Fraud Detection

RISE Internship — Machine Learning & AI

---

Problem Statement

Financial institutions and e-commerce platforms face increasing fraudulent transactions resulting in major financial losses. Rule-based systems fail to detect evolving fraud patterns. This project builds an ML model that identifies suspicious transactions in real time.

Objective

Build a machine learning model that detects fraudulent transactions based on transaction behavior and historical data.

---

Project Structure

fraud_detection/
├── fraud_detection.py          ← Main script (all steps end-to-end)
├── README.md                   ← This file
└── outputs/
    ├── 01_eda_overview.png             ← Class balance, amount dist, fraud by hour
    ├── 02_feature_distributions.png   ← Feature distributions by class
    ├── 03_model_comparison.png        ← Precision, Recall, F1, AUC bar chart
    ├── 04_confusion_matrices.png      ← Confusion matrix for all 4 models
    ├── 05_roc_curves.png              ← ROC curves comparison
    ├── 06_precision_recall_curve.png  ← P-R curve for best model
    ├── 07_feature_importance.png      ← Top 15 features (Random Forest)
    ├── 08_fraud_patterns_correlation.png ← Fraud vs Legit correlation heatmaps
    └── suspicious_transactions.csv   ← Flagged high-risk transactions

---

Tools & Libraries

Tool	Purpose
Python	Core programming language
NumPy	Numerical operations
Pandas	Data manipulation
Matplotlib	Plotting & visualization
Seaborn	Statistical visualizations
Scikit-learn	ML models & evaluation
Jupyter Notebook	Interactive development

---

Pipeline Steps

Step 1 — Dataset Ingestion

• Synthetic dataset of 10,000 transactions (9,500 legit + 500 fraud)
• Features: Time, Amount, V1–V5 (behavioral features), Hour, MerchantCategory
• Realistic 5% fraud rate to simulate class imbalance

Step 2 — EDA (Exploratory Data Analysis)

• Class distribution visualization
• Transaction amount comparison (fraud vs legit)
• Fraud frequency by hour of day
• Feature distribution overlays

Step 3 — Data Cleaning & Preprocessing

• Zero missing values verified
• One-hot encoding for categorical feature (MerchantCategory)
• StandardScaler normalization on Amount and Time

Step 4 — Handling Class Imbalance

• Train/test split (80/20) with stratification — BEFORE resampling
• Oversampling (resample) applied only to training data
• Balanced 50/50 split for training

Step 5 — Feature Engineering

• V1×V2 interaction feature
• V4×V5 interaction feature
• Amount/V1 ratio
• V-feature sum and standard deviation

Step 6 — Model Training

Four classifiers trained and compared:

• Logistic Regression
• Decision Tree (max_depth=8)
• Random Forest (100 estimators)
• Gradient Boosting (100 estimators)

Step 7 — Model Evaluation

Metrics used: Precision, Recall, F1-Score, ROC-AUC

Model	Precision	Recall	F1	ROC-AUC
Logistic Regression	0.840	1.000	0.913	1.000
Decision Tree	0.825	0.990	0.900	0.991
Random Forest	1.000	0.950	0.974	1.000
Gradient Boosting	0.952	1.000	0.976	1.000

Best Model: Gradient Boosting (F1 = 0.9756)

Step 8 — Fraud Pattern Visualization

• Correlation heatmaps (fraud vs legit)
• ROC curves for all models
• Precision-Recall curve
• Feature importance chart

Step 9 — Suspicious Transaction Prediction

• Fraud probability score generated for all test transactions
• Transactions with probability > 0.70 flagged as suspicious
• Exported to CSV for business action

---

Key Results

• 99 out of 99 suspicious-flagged transactions were confirmed fraud
• Gradient Boosting achieved perfect recall — no fraud transactions missed
• Top fraud indicators: V2 (high positive), V1 (high negative), V5 (high negative)

---

RISE 4.0 Internship | Tamizhan Skills