Interpretable Machine Learning for Fraud Detection in Large-Scale Financial Transactions

Abstract

This project presents a research-oriented and interpretable machine learning framework for detecting fraudulent financial transactions in large-scale banking data. Using a dataset of approximately 18 million transactions, the system emphasizes behavioral feature engineering, probabilistic risk estimation, and post-hoc analytical evaluation. An interactive experimental dashboard is provided to support controlled transaction testing, qualitative error analysis, and interpretability-driven inspection. This project is designed as an academic research artifact, suitable for MSc applications and future peer-reviewed publication.

---

1. Research Motivation

Fraud detection in financial systems presents multiple real-world and research challenges:

• Extreme class imbalance: Fraud cases represent less than 0.2% of the data.
• High cost of false negatives: Leading to direct financial loss.
• Regulatory demand for interpretability: Modern systems require explainable decision-making.

This work focuses on:

• Modeling transactional behavior rather than customer identity.
• Combining statistical learning with domain-driven financial indicators.
• Producing transparent risk outputs instead of opaque binary decisions.

---

2. Dataset Description

• Source: Financial Transactions Dataset (AIML – Synthetic & Anonymized).
• Scale: ~18,000,000 transactions.
• Fraud Ratio: ~0.17%.
• Transaction Types: TRANSFER, CASH_OUT, PAYMENT, DEBIT, CASH_IN.
• Ethical Note: Contains no Personal Identifiable Information (PII).

---

3. System Overview & Interactive Evaluation

3.1 Transaction Input Interface

The system provides an interactive interface that allows manual transaction simulation, enabling controlled experimentation with transaction attributes.

3.2 Prediction Output & Risk Interpretation

For each evaluated transaction, the system outputs a Fraud Probability, Aggregated Risk Score, and Behavioral Alerts.

---

4. Exploratory Data & Model Behavior Analysis

4.1 Fraud Probability Distribution

The distribution highlights how the model allocates probability mass under extreme class imbalance.

4.2 Transaction Amount vs Fraud Label

A log-scaled comparison demonstrates separation trends between legitimate and fraudulent transactions.

4.3 Transaction Type Analysis

Analysis of fraud incidence across different operation types (notably higher in TRANSFER and CASH_OUT).

4.4 Feature Correlation Analysis

Highlights relationships between transaction attributes and engineered behavioral features.

---

5. Model Behavior & Error Analysis

5.1 Amount vs Predicted Fraud Probability

This visualization reveals the concentration of predicted risk in specific transaction regimes.

5.2 Top Suspicious Transactions

A ranked view of high-risk transactions supports manual audit and qualitative error analysis.

---

6. Methodology Summary

• Model: Logistic Regression (Interpretable baseline).
• Preprocessing: Standardization and One-hot encoding.
• Feature Engineering:
  • Balance differentials.
  • Amount-to-balance ratios.
  • Behavioral risk flags.

---

7. Research Value

This project demonstrates applied machine learning in a realistic financial setting, focusing on:

• Handling highly imbalanced datasets.
• Behavioral feature engineering grounded in domain logic.
• Reproducible and inspectable experimental analysis.

Suitable As: MSc research portfolio material or a foundation for peer-reviewed research.

---

8. Limitations & Future Work

• Temporal modeling: Using sequence-based approaches (LSTM / Transformers).
• Explainable Ensembles: Integrating SHAP or LIME with Gradient Boosting.
• Real-time deployment: Implementing streaming inference pipelines.

---

Author

Mariam Zakaria MSc Applicant — Machine Learning & Data Science Research Interests: Fraud Detection, Interpretable Machine Learning, Applied AI Systems.