Fraud Detection System for Financial & E-commerce Transactions

An end-to-end, production-ready fraud detection platform designed to identify suspicious transactions in real time. The system transforms raw transactional and behavioral data into fraud signals, trains and tracks machine learning models, and exposes predictions through a scalable API.

---

Table of Contents

• Project Overview
• Business Context
• Fraud Detection Business Understanding
• Objectives
• Dataset Overview
• Project Structure
• Architecture
• Modeling Approach
• MLOps & Engineering Practices
• Setup & Installation
• Running the Project
• API Usage
• Technologies Used
• Author

---

Project Overview

This project implements a Fraud Detection System for financial and e-commerce transactions. It is designed to detect fraudulent behavior in environments where transaction volume is high, patterns evolve quickly, and false positives are costly.

The system covers the full lifecycle:

• Data ingestion and cleaning
• Feature engineering (temporal, behavioral, transactional, geolocation)
• Class imbalance handling (SMOTE / undersampling)
• Model training, tuning, and evaluation
• Explainability with SHAP and feature importance
• Experiment tracking and reproducibility
• Real-time fraud prediction via FastAPI
• Containerized deployment using Docker

---

Business Context

Financial institutions and e-commerce platforms face continuous losses due to fraudulent transactions. Fraud patterns change rapidly, making rule-based systems brittle and hard to maintain.

This project supports:

• Real-time fraud screening
• Risk-based transaction approval
• Fraud monitoring and investigation
• Reduction of false positives

The output is a fraud probability score that can be integrated into transaction decision pipelines.

---

Fraud Detection Business Understanding

Nature of Fraud Problems

Fraud detection differs from traditional classification tasks:

• Fraud cases are rare (high class imbalance)
• Fraud patterns evolve over time
• False positives directly impact customer experience
• Explainability is required for investigations and audits

This system is designed with these constraints in mind.

Class Imbalance Strategy

Fraud labels are highly imbalanced. To address this, the project supports:

• SMOTE and controlled oversampling
• Undersampling of majority class
• Threshold tuning based on business cost
• Precision-Recall focused evaluation (AUC-PR, F1)

Model Interpretability

Fraud predictions must be explainable for:

• Internal fraud analysts
• Customer dispute resolution
• Regulatory and audit requirements

This project integrates:

• SHAP value analysis
• Global and local feature importance
• Human-readable explainability reports

---

Objectives

• Detect fraudulent transactions accurately
• Engineer robust fraud-specific features
• Handle extreme class imbalance safely
• Compare interpretable and complex models
• Explain model predictions
• Serve predictions through a scalable API
• Ensure reproducibility and auditability

---

Dataset Overview

Sources:

• E-commerce transaction logs
• Banking / credit card transaction datasets
• IP geolocation reference data

Key fields include:

Column	Description
TransactionId	Unique transaction identifier
CustomerId	Unique customer identifier
Amount	Transaction amount
ChannelId	Platform channel
TransactionStartTime	Timestamp
IPAddress	Client IP address
FraudResult	Fraud label (0 = legit, 1 = fraud)

Derived features include:

• Time-based signals
• Velocity and frequency features
• Amount-based aggregations
• Country and geolocation risk indicators

---

Project Structure

fraud-detection/
│
├── config/                          # YAML configuration files for system behavior and experiments
│
├── data/                             # Storage for all dataset versions
│   ├── raw/                          # Original datasets (never modified)
│   ├── interim/                      # Cleaned but not fully transformed data
│   ├── processed/                    # Final model-ready datasets for training and inference
│
├── notebooks/                        # Jupyter notebooks for EDA, feature engineering, modeling
│
├── src/
│   └── fraud_detection/               # Main Python package
│       ├── __init__.py                # Marks fraud_detection as a package
│
│       ├── analysis/                  # Exploratory data analysis tools
│
│       ├── core/                      # Global core functionality and configuration
│
│       ├── data/                      # Data ingestion and preprocessing
│
│       ├── explainability/            # Model explainability and interpretation
│
│       ├── features/                  # Feature engineering modules
│
│       ├── models/                    # Model training, evaluation, and persistence
│
│       ├── utils/                      # Shared helper functions
│
│       └── viz/                        # Visualization utilities
│
├── tests/                                # Automated testing
│
│
├── scripts/                              # CLI & automation entry points
│
├── mlruns/                               # MLflow experiment tracking
├── dvc.yaml                              # DVC pipeline definition
├── params.yaml                           # Global experiment parameters
├── requirements.txt                     # Runtime dependencies
├── pyproject.toml                       # Build & packaging configuration
└── README.md                            # Project overview & usage



## Architecture
```textRaw Data
   ↓
Data Loading & Validation
   ↓
IP Geolocation Enrichment
   ↓
Feature Engineering Pipeline
   ↓
Model Training & Hyperparameter Tuning
   ↓
Experiment Tracking (MLflow)
   ↓
Explainability (SHAP + Feature Importance)
   ↓
FastAPI Inference Layer
   ↓
Streamlit Risk Dashboard

Setup & Installation

Clone the repository:

git clone https://github.com/<useraname>/fraud_detection.git
cd fraud_detection

Create and activate a virtual environment:

python -m venv .venv
source .venv/bin/activate

Install dependencies:

• pyproject run

pip install -r requirements.txt

• pyproject run

pip install -e .

Running the Project

Run the full pipeline (DVC)

dvc repro

Run Dashboard

streamlit run dashboard/app.py

Dashboard Demo

Author

Tibebu Kaleb | ML/AI Engineer

🌐 Connect