Fraud Detection using Machine Learning

An end-to-end fraud detection project that develops and compares statistical and machine learning models for identifying fraudulent financial transactions using Python.

---

Overview

Fraud detection is a highly imbalanced binary classification problem where fraudulent transactions represent only a small fraction of all observations. This project develops a complete analytical pipeline, from data preprocessing and exploratory analysis to predictive modelling and performance evaluation.

Both classical statistical models and machine learning techniques are implemented and compared to assess their effectiveness in identifying fraudulent transactions.

Project Summary

Attribute	Value
Domain	Fraud Detection
Dataset	Synthetic Financial Transactions
Observations	12,000
Features	24
Fraud Rate	2.13%
Models	Linear Regression, Logistic Regression, MLP, Autoencoder
Language	Python

Objectives

• Develop an end-to-end fraud detection pipeline.
• Compare statistical and machine learning approaches for fraud detection.
• Evaluate predictive performance using metrics suitable for imbalanced classification.
• Identify the main factors associated with fraudulent transactions.

Dataset

The analysis uses a synthetic financial transactions dataset containing:

• 12,000 observations
• 24 variables
• Fraud rate of 2.13%

The dataset includes customer characteristics, transaction attributes, behavioural indicators and a binary fraud label.

Methodology

The project follows the workflow below:

1. Data preprocessing
2. Missing value imputation
3. Exploratory Data Analysis (EDA)
4. Feature engineering
5. Model development
6. Model evaluation
7. Performance comparison

Models

The following predictive models were implemented:

• Linear Regression
• Logistic Regression
• Multilayer Perceptron (MLP)
• Autoencoder

Evaluation Metrics

Model performance was assessed using:

• Precision
• Recall
• F1-score
• ROC-AUC

Special emphasis was placed on metrics appropriate for highly imbalanced classification problems.

Results

The analysis identified several important fraud indicators, including customer risk score, transaction amount ratio, foreign transactions, night transactions and high-risk merchants.

The supervised models achieved excellent predictive performance, while the autoencoder demonstrated the applicability of unsupervised anomaly detection techniques.

The following figures present selected visualizations from the exploratory analysis and model evaluation.

Technologies

• Python
• pandas
• NumPy
• scikit-learn
• Matplotlib

Repository Structure

fraud-detection-python/
│
├── data/
├── results/
├── report/
└── README.md

Repository Contents

Folder	Description
data	Dataset used in the analysis
results	Figures and model outputs
report	Final project report

Key Visualizations

Correlation Heatmap

ROC Curve

Confusion Matrix

Author

Konstantinos Papavrontos

MSc Candidate in Business Economics with Analytics
Athens University of Economics and Business