Bank Transaction Fraud Detection – Machine Learning Project

Overview

This project explores a fraud detection problem using bank transaction data. The objective is not to make hard fraud decisions, but to build a risk-oriented predictive model that estimates the probability of a transaction being fraudulent, supporting decision-making under uncertainty.

The project covers the full data science pipeline:

• Exploratory Data Analysis (EDA)
• Feature engineering
• Handling strong class imbalance
• Model training and evaluation
• Feature importance and explainability (SHAP)
• Risk-based prediction

---

Dataset

The dataset was obtained from Kaggle: Bank Transaction Fraud Detection

Due to size and licensing constraints, the full dataset is not included in this repository. A representative sample is used for preview and reproducibility. The complete dataset can be accessed directly on Kaggle.

---

Problem Characteristics

• Highly imbalanced dataset (~5% fraud)
• Fraud patterns are subtle and overlap strongly with non-fraud transactions
• Standard accuracy metrics are misleading
• Focus on ROC-AUC, relative risk, and probabilistic outputs

---

Feature Engineering

Key feature engineering steps include:

• Removal of non-informative identifiers (UUIDs, IDs, timestamps)
• Transformation of repeated identifiers into frequency-based features
• Selection of financially meaningful numerical variables

Final core features used for modeling:

• Transaction Amount
• Account Balance
• Customer Age

---

Exploratory Analysis

Several exploratory analyses were performed, including:

• Fraud distribution and class imbalance
• Fraud rates across bank branches
• Relative fraud risk by bank (normalized by global average)

---

Modeling Approach

A Random Forest Classifier was chosen due to:

• Non-linear decision boundaries
• Robustness to feature scaling
• Interpretability via feature importance and SHAP

Class imbalance was handled using:

• Stratified train-test split
• Class weighting (class_weight='balanced')

Model evaluation focused on:

• ROC-AUC
• Feature ablation (AUC drop)

---

Feature Importance & Ablation Study

Two complementary approaches were used:

Impurity-Based Importance

Random Forest feature importances highlighted the dominance of numerical variables.

AUC-Based Feature Ablation

Each feature was removed individually, retraining the model to measure the real impact on predictive performance.

This confirmed that:

• Transaction Amount
• Account Balance
• Age

are the most influential features in fraud prediction.

---

Model Explainability (SHAP)

SHAP values were used to analyze how features influence predictions at an individual level.

Key observations:

• Most SHAP values are concentrated around the base value
• The model rarely receives strong evidence to push predictions toward fraud
• This behavior is consistent with strong class imbalance and subtle fraud patterns

---

Risk-Oriented Prediction

Instead of outputting a binary decision, the final model produces a fraud risk score:

Probability that a transaction is fraudulent

This allows:

• Flexible thresholding
• Integration with business rules
• Manual review prioritization

#Example usage:

predict_fraud_risk(model, amount=50000, balance=20000, age=60)

---

Limitations

• Dataset limited to a single country and month
• Fraud patterns may evolve over time
• Model performance constrained by data imbalance and feature overlap

---

Future Work

• Cost-sensitive learning
• Time-based validation
• Threshold optimization based on business cost
• Integration with real-time alert systems

---

Technologies Used

• Python
• Pandas / NumPy
• Scikit-learn
• SHAP
• Matplotlib / Seaborn

---

Author

Bastián B. BSc in Mathematics Aspiring Data Analyst / Data Scientist