ML Pipeline for Imbalanced Data Classification

---

Project Description

This project focuses on detecting fraudulent credit card transactions using different machine learning models.
It handles highly imbalanced data and provides an end-to-end pipeline for preprocessing, sampling, model training, and evaluation.
The workflow includes data scaling, oversampling/undersampling, hyperparameter tuning, and evaluation using metrics like F1-score and Precision-Recall AUC.

---

Implemented Models & Sampling Options

Models selection (configurable via the config files):

• Logistic Regression (with configurable hyperparameters and optional Grid Search)
• Random Forest Classifier (with configurable hyperparameters and optional Randomized Search)
• Voting Classifier (combining Logistic Regression and Random Forest)

Sampling & Scaling options:

• Robust scaler
• Standard Scaler
• Oversampling: SMOTE, RandomOverSampler
• Undersampling: RandomUnderSampler, NearMiss
• Hybrid: SMOTETomek

Evaluation options:

• f1-Score
• PR-AUC
• Confusion Matrix
• Best threshold for the model using PR Curve
• Custom threshold

All models and sampling methods can be configured through YAML configuration files, allowing flexible experimentation.

---

Repository Structure

credit_fraud_detection/
│
├─ config/
│ ├─ config.yml
│ ├─ model_config.yml
│ └─ eval_config.yaml
│
├── data/
│   ├── train.csv
│   ├── val.csv
│   └── test.csv
│
├─ models/
│  ├──
│  └── 
├── cradit_fraud_utils_helper.py
├── credit_fraud_train.py
├── credit_fraud_utils_data.py
├── credit_fraud_utils_eval.py
├── EDA.ipynb
├── model_test.py
└─ README.md

---

Files Description

• config/

  • config.yaml : Contains data paths and data preprocessing & balancing configuration settings
  • eval_config.yaml : Contains evaluation metrics configuration
  • model_config.yaml : Contains model selection, and configuration settings specific to each trained models which is :
    • Hyperparameters
    • Grid search
    • randomized search

• data/

  • train.csv: The training dataset.
  • val.csv: The validation dataset.
  • test.csv: The test dataset.

• models/ : Directory for storing trained models, plots and the config files for the model.

• credit_fraud_train.py : Script to train models with optional sampling, scaling, and hyperparameter tuning. Saves trained models and thresholds.

• credit_fraud_utils-data.py : Contains functions and utilities for data preprocessing and Loading.

• credit_fraud_utils_helper.py : Helper functions for loading models, configuration files.

• credit_fraud_utils_eval.py : Evaluation functions including F1-score, precision-recall curves, confusion matrices, and threshold optimization.

• model_test.py: Script for evaluating the trained models on test data

• eda.ipynb: Jupyter Notebook for Exploratory Data Analysis (EDA) on the training dataset.

---

Results

Models performance on test dataset

Models performance against random classifier