Credit Card Fraud Detection: A Complete ML Pipeline

A Comparative Study of Resampling and Modeling Strategies

An end-to-end machine learning pipeline for detecting fraudulent credit card transactions, leveraging advanced resampling techniques and ensemble modeling to tackle class imbalance and improve predictive performance.

---

Overview

This project addresses credit card fraud detection using supervised and unsupervised machine learning on a highly imbalanced dataset (~0.18% fraud). It builds a complete ML pipeline covering preprocessing, exploratory analysis, feature engineering, and advanced resampling.

To tackle class imbalance and improve performance, we applied techniques like random undersampling, SMOTE, and a custom KMeans-based method, paired with models such as logistic regression, random forest, KNN, neural networks, and an ensemble voting classifier.

The best-performing model was selected through threshold tuning and F1-score optimization, followed by targeted feature engineering to further boost performance. The pipeline is modular, reproducible, and CLI-configurable—ready for real-world deployment.

---

Dataset

• Source: Kaggle - Credit Card Fraud Detection

• Total Transactions: 284,807

• Fraudulent Cases: 492

• Features:

  • V1 to V28 (PCA-transformed)
  • Amount, Time
  • Class (target: 1 = fraud, 0 = legitimate)

---

Key Highlights

• Implemented six resampling techniques to address class imbalance:

  • baseline (cost-sensetive)
  • Random Undersampling
  • NearMiss
  • Custom KMeans-based undersampling
  • Random Oversampling
  • SMOTE

• Trained and compared five machine learning models:

  • Logistic Regression
  • Random Forest
  • K-Nearest Neighbors
  • Neural Network (MLPClassifier)
  • Ensemble Voting Classifier combining top performers

• Custom KMeans-based undersampler developed to improve representation of the majority class while preserving data structure

• Feature engineering to enhance model performance, supported by feature importance analysis to identify influential variables

• Applied threshold tuning based on F1-score to better balance precision and recall

• Modular, CLI-driven architecture for reproducible and configurable experimentation

• Professional final report and cleanly structured Jupyter notebooks for analysis, experimentation, and reproducibility

• Designed for easy extensibility to test new resampling or modeling approaches

---

Model Performance Overview

Model	F1 Score	Precision	Recall	PR-AUC
Neural Network (threshold=0.606)	85.25%	90.70%	80.41%	85.70%
Random Forest (threshold=0.364)	84.69%	83.84%	85.57%	85.01%
K-Nearest Neighbors (threshold=0.414)	87.05%	87.50%	86.60%	82.81%
Voting Classifier (threshold=0.283)	86.73%	85.86%	87.63%	87.73%

---

📄 Full Report

The report includes:

• Motivation and problem framing
• Exploratory data analysis insights and class distribution
• Feature engineering strategies
• In-depth resampling method comparison
• Model evaluation and threshold tuning
• Metric interpretation and real-world considerations

👉 View Full Report (PDF)

---

Project Structure

Credit-Card-Fraud-Detection/
├── notebooks/
│   └── credit_fraud.ipynb             # Exploratory Data Analysis (EDA) notebook
│
├── summary/
│   └── Project_Report.pdf             # Final project report document
│
├── scripts/
│   ├── credit_fraud_train.py          # Script for training models
│   ├── credit_fraud_test.py           # Script for testing models
│   │
│   ├── credit_fraud_utils_data.py     # Data loading, cleaning, and splitting functions
│   ├── credit_fraud_utils_eval.py     # Evaluation metric functions
│   ├── kmeans_undersampler.py         # Custom KMeans undersampling implementation
│   └── cli_args.py                    # Command-line argument parser
│
├── saved_models/                      # Folder containing all saved models
│   ├── knn_final_model.pkl
│   ├── nn_final_model.pkl
│   ├── rf_feature_importance_model.pkl
│   ├── rf_final_model.pkl
│   └── vc_final_model.pkl
│
├── README.md                         # Project overview and usage instructions
└── requirements.txt                  # Python package dependencies list

---

Installation

Clone the repository:

git clone https://github.com/your-username/Credit-Card-Fraud-Detection.git
cd Credit-Card-Fraud-Detection

Install dependencies:

pip install -r requirements.txt

Python 3.9 or later is required for full compatibility.

---

Command-Line Interface (CLI Options)

This project supports flexible configuration via command-line arguments using argparse. Below is a breakdown of available CLI options by category.

File Paths

Argument	Description
--train-dir	Path to the training dataset CSV (default: ../data/split/trainval.csv)
--test-dir	Path to the test dataset CSV (default: ../data/split/test.csv)
--save-dir	Directory to save trained model artifacts (default: ../saved_models/)
--load-dir	Path to load a saved model for evaluation (default: ../saved_models/final_model.pkl)
--model-name	Name used when saving the model file (e.g., model_v1.pkl)

Preprocessing Options

Argument	Description
--scaling	Feature scaling method: standard, minmax, robust, or raw
--balance	Class balancing method: none, undersample, nearmiss, kmeans, cluster_centroids, oversample, smote
--n-neighbors	Number of neighbors for SMOTE or NearMiss (default: 5)
--nearmiss-version	Version of NearMiss to use (1, 2, or 3)
--n-clusters	Number of clusters for KMeans undersampling (default: 500)
--use-pca	Flag to apply PCA before training (default: False)
--n-components	Number of PCA components to retain (default: 18)

Model Selection

Argument	Description
--model	Model to train: LR (Logistic Regression), NN (MLP), RF, VC, or KNN

Model Hyperparameters

K-Nearest Neighbors (KNN)

Argument	Description
--knn-n-neighbors	Number of neighbors to use (default: 7)
--knn-weights	Weight strategy: uniform or distance

Logistic Regression (LR)

Argument	Description
--lr-c	Inverse of regularization strength (default: 1)
--lr-max-iter	Maximum iterations (default: 10000)

Random Forest (RF)

Argument	Description
--rf-n-estimators	Number of trees (default: 100)
--rf-max-depth	Maximum tree depth (default: None)

Neural Network (NN / MLP)

Argument	Description
--nn-hidden-layers	Comma-separated hidden layer sizes (e.g., 128,64)
--nn-max-iter	Maximum training iterations (default: 3000)
--nn-activation	Activation function: relu, tanh, or logistic
--nn-lr	Learning rate (default: 0.001)
--nn-alpha	L2 regularization parameter (default: 0.001)

Additional Options

Argument	Description
--grid-search	Enable GridSearchCV for hyperparameter tuning
--cost-sensitive	Use cost-sensitive learning via class weighting

Example Usage

Train a neural network with SMOTE and PCA:

python scripts/credit_fraud_train.py \
    --train-dir data/split/trainval.csv \
    --model NN \
    --balance smote \
    --scaling standard \
    --use-pca \
    --model-name smote_nn_pca

---

Tools and Technologies

• Python 3.9+
• scikit-learn
• imbalanced-learn
• pandas, numpy
• matplotlib, seaborn
• argparse
• t-SNE (sklearn.manifold)
• Jupyter Notebook

---

Contact

Author: Taher Alabbar
Email: t.alabbar.ca@gmail.com, LinkedIn