Credit Card Fraud Detection API

This project is a complete machine learning pipeline built to detect fraudulent credit card transactions, along with a deployed API to serve predictions in real-time.

The goal was not just to train a model, but to understand the problem deeply and build something end-to-end — from data preprocessing to deployment.

---

Problem Context

Credit card fraud detection is a classic imbalanced classification problem.

• Fraud cases are extremely rare (~0.17%)
• Most transactions are legitimate
• This makes accuracy a misleading metric

The dataset used is publicly available on Kaggle: mlg-ulb/creditcardfraud

Important note:

• Features V1–V28 are anonymized (PCA transformed)
• Only Amount and Time are interpretable

---

Approach

1. Data Preprocessing

• Removed duplicates
• Checked for missing values
• Performed basic statistical analysis
• Observed extreme class imbalance

---

2. Exploratory Data Analysis (EDA)

• Visualized class distribution
• Analyzed transaction amounts vs fraud
• Checked time-based patterns
• Generated correlation heatmap

Key insight:

Fraud detection requires focusing on minority class behavior, not overall accuracy.

---

3. Handling Class Imbalance

Used SMOTE (Synthetic Minority Oversampling Technique)

Why?

• Original fraud samples were too few
• Model would otherwise ignore fraud cases
• SMOTE helps balance the dataset artificially

---

4. Feature Scaling

• Scaled Amount and Time using StandardScaler
• Other features already transformed (PCA)

---

5. Model Training

Trained multiple models instead of jumping to one:

• Logistic Regression → baseline
• Random Forest → main model
• ExtraTrees + HistGradientBoosting → additional comparison

Why multiple models?

To compare performance and understand trade-offs instead of blindly choosing one.

---

6. Model Evaluation

Used:

• Precision
• Recall
• F1-score (especially for fraud class)

Why not accuracy?

Because predicting "not fraud" always would still give ~99% accuracy.

---

7. Threshold Tuning

Instead of relying on default 0.5:

• Tested multiple thresholds (0.5 → 0.95)

Why?

• Fraud detection requires balancing:
  • catching fraud (recall)
  • avoiding false alarms (precision)

---

8. Final Model Selection

Random Forest was selected

Reason:

• Best balance between precision and recall
• Strong performance on minority class
• Fewer false positives compared to others

---

API Design

Built using FastAPI and deployed on Render.

Live API:

https://credit-card-fraud-detection-api-bx3x.onrender.com

---

How Prediction Works

User provides:

{
  "amount": 50,
  "time": 10000
}

---

Important Design Decision

Why only Amount and Time as input?

Because:

• Other features are anonymized (PCA)
• Their real-world meaning is unknown
• Asking users for them is not practical

So:

Default (mean) values are used for remaining features

This is a practical compromise, not a perfect solution.

---

Running Locally

git clone https://github.com/rohitnath-dev/credit-card-fraud-detection-api.git
cd credit-card-fraud-detection-api

pip install -r requirements.txt

uvicorn app.main:app --reload

API Documentation

Open the API docs here:
http://127.0.0.1:8000/docs

---

Limitations

• Dataset is anonymized → limited interpretability
• Input features are incomplete (only 2 real inputs)
• Model is not trained on real-world production data
• SMOTE creates synthetic samples (not real fraud cases)

---

Future Improvements

• Use real-world transaction features
• Better feature engineering
• Advanced models (XGBoost, LightGBM)
• Real-time streaming system
• Frontend integration

---

Final Note

This project focuses on understanding the full ML pipeline, not just model training.

data → model → evaluation → API → deployment

Everything is built and connected end-to-end.