Autoencoder-Based Credit Card Fraud Detection

---

This project implements an unsupervised deep learning pipeline using an autoencoder to detect fraudulent transactions in credit card data. It follows a robust ML workflow with modular design, EDA, preprocessing, 5-fold CV training, anomaly detection using reconstruction loss, and SHAP-based interpretability.

Project Structure

---

creditcard_autoencoder/
│   main.py
│   README.md
│
├─── /data
│       creditcard.csv
│
├─── /src
│       config.py
│       eda.py
│       evaluate.py
│       interpret.py
│       model.py
│       preprocess.py
│       train.py
│
└─── /notebooks
        exploration.ipynb

---

Features Used

From the original dataset, the following columns are selected and processed:

• PCA-Transformed Features: V1 to V28
• Original Features: Time, Amount
• Target: Class (0 = Normal, 1 = Fraudulent)

Transactions are normalized and fraud is detected by analyzing reconstruction error using an unsupervised autoencoder.

---

Workflow

1. Data Preprocessing (preprocess.py)

• Time and Amount are scaled with StandardScaler.
• Data is splot into train/test with stratification on class.
• Only normal transactions are used to train the autoencoder.

2. Exploratory Data Analysis (eda.py)

• Class imbalance visualization.
• Feature distribution histograms.
• Correlation heatmaps.
• PCA/t-SNE visualization of transaction clusters.

3. Autoencoder Model (model.py)

• Trained to minimize reconstruction loss (MSE).
• Tuned using parameters like encoding dimension, batch size, epochs, etc.

4. Training with Cross-Validation (train.py)

• Used 5-fold Stratified K-Fold Cross Validation.
• Trains on Class 0 (normal) data only.
• Final model selected and evaluated on the full test set.

5. Evaluation (evaluate.py)

• Predicts reconstruction error for test transactions.
• Flags frauds based on a threshold.
• Metrics: Confusion Matrix, Precision, Recall, F1, AUC

6. Explainability (interpret.py)

• Uses SHAP to identify which features contribute most to reconstruction error.
• Helps interpret model behavior for both normal and fraud predictions.

---

Requirements

Installing dependencies:

pip install -r requirements.txt

---

Running the project

python main.py

The pipeline will:

1. Load and explore the data
2. Preprocess features and scale them
3. Train teh autoencoder with K-Fold Cross Validation
4. Evaluate on test set with anomaly scoring
5. Generate interpretability plots with SHAP

---

Current Output

Results:

Best parameters:

---

Dataset

The dataset used sources from Machine Learning Group - ULB Credit Card Fraud Detection Kaggle dataset website.

---

Future improvements

• Compare autoencoder with isolation forests, One-Class SVM
• Hyperparameter tuning via KerasTuner
• Deploy API for fraud scoring
• Real-time stream integration with Kafka or Spark
• Web dashboard for fraud alerts

---

License

This project is licensed under the MIT License.