Airline-Payment-Fraud-XGBoost

Airline Payment Fraud XGBoost Model

I built this XGBoost model to estimate the probability that an airline payment transaction is fraudulent. This model is part of my Airline Payment Fraud Decision Intelligence System, an end-to-end data science project that goes beyond fraud classification. I use the predicted fraud probability as a risk score, analyze the model's behavior with XAI, and then apply a decision threshold to support operational recommendations.

What I Built?

I designed this project around a simple workflow:

Transaction Data ⇰ Fraud Risk Prediction ⇰ Threshold Analysis ⇰ XAI ⇰ Decision Recommendation

The XGBoost model is responsible for the prediction stage. It produces a fraud probability for each transaction rather than directly making a business decision. I then use a decision layer to translate that probability into a more meaningful operational recommendation.

This distinction is important: predicting fraud risk and deciding what action to take are two separate steps.

Model Details

• Algorithm: XGBoost
• Task: Binary classification
• Target: is_fraud
• Output: Fraud probability
• Model artifact: airline_fraud_xgboost.joblib

I trained the model using structured airline payment transaction data and selected the final feature set through the project's data preparation and feature engineering workflow. I evaluated the model on a held-out test set and used threshold analysis to understand the trade-off between detecting fraudulent transactions and unnecessarily interrupting legitimate transactions.

Dataset

The dataset I used for this model contains approximately 3 million airline payment transaction records covering transaction, payment, account, geographic, behavioral, and session-related information.

During the project, I treated identifier-like and potentially sensitive fields carefully and did not expose raw values from fields such as email addresses, IP addresses, payment identifiers, card information, or device identifiers in the public project.

The dataset has its own licensing and usage terms. The MIT license shown on this model repository applies to the materials I publish here and does not automatically apply to the underlying dataset.

Model Evaluation

I evaluated the final model using metrics appropriate for a fraud-risk classification problem.

I also evaluated different probability thresholds rather than treating the default 0.50 classification threshold as automatically appropriate. For this project, I selected a 0.91 fraud-probability threshold for the decision layer. This threshold is specific to the project's analytical objective and should not be interpreted as a universal threshold for airline fraud detection.

Decision Layer

I deliberately separated the model's prediction from the final operational recommendation.

The model produces: Fraud Probability ⇰ Risk Assessment

The decision layer then uses the selected threshold and the transaction's position relative to that threshold to support recommendations such as:

• Low risk: Approve
• High risk: Flag
• Near threshold: Additional verification or manual review may be appropriate

This approach allows the model to function as part of a broader decision-support system rather than treating a classification output as an automatic business decision.

Explainability

I used SHAP (SHapley Additive exPlanations) to understand how the model arrives at its predictions.

I used XAI at two levels:

1. Global interpretation for understanding which features generally influence the model's predictions.
2. Individual interpretation for examining why a particular transaction received a specific fraud-risk score.

These explanations help make the model's behavior easier to inspect and communicate.

However, SHAP explanations describe model behavior, not causality. A feature contributing strongly to a prediction does not necessarily mean that the feature causes fraud.

Intended Use

I built this model for:

• Airline payment fraud-risk analysis.
• Transaction-level fraud probability estimation.
• Demonstrating an end-to-end fraud decision-support workflow.
• Exploring interpretable machine-learning approaches for tabular fraud data.
• Portfolio, research, and educational purposes.

The model is intended to support decision-making rather than replace a complete fraud-management system.

Limitations

There are several limitations to keep in mind.

1. Dataset limitations: I trained the model on a specific historical dataset, so its performance may not generalize to another airline, payment provider, geography, or future time period.
2. Changing fraud patterns: Fraud behavior can change over time. New attack patterns, payment methods, customer behavior, or operational processes can introduce model drift.
3. Threshold limitations: The 0.91 threshold is specific to this project. A real organization would need to select its threshold using current fraud prevalence, operational capacity, customer impact, and the actual costs of false positives and false negatives.
4. Probability interpretation: The model's output is a predicted probability, but I have not treated that probability as a guarantee of perfect real-world calibration.
5. Business-cost assumptions: The dataset does not provide the complete monetary cost of fraud, false positives, customer friction, or manual review. Therefore, I treat those factors as business considerations rather than observed financial outcomes.

Responsible Use

I would not use this model as the sole basis for automatically approving or rejecting real-world payment transactions without additional validation and operational controls.

Before production use, an organization should independently evaluate:

• Performance on current transaction data.
• Probability calibration.
• Threshold selection using real business costs.
• False-positive and false-negative consequences.
• Data quality and feature availability.
• Model drift over time.
• Population and geographic differences.
• Privacy, security, and regulatory requirements.

The model should be treated as one component of a broader fraud-risk management process.

Model Artifact

I saved the trained model as: airline_fraud_xgboost.joblib

The artifact contains the trained XGBoost model used by the project's inference workflow. I use this model artifact in the Streamlit app that accompanies the project.

Project Resources

1. GitHub Repository: The complete project source code, notebooks, application code, and project structure are available here: Airline Payment Fraud Decision Intelligence System

2. Live Fraud Detection App: Airline Fraud Detection App

3. Project Walkthrough: I also documented the complete project journey, from the business problem through modeling, XAI, decision-making, and deployment: How I Built an Airline Payment Fraud Detection System with Machine Learning and Explainable AI?

License

I released the materials in this model repository under the MIT License.

The underlying training dataset may have separate licensing and usage conditions.

Disclaimer

This is a portfolio and research-oriented project. I have not validated this model for production use in a live airline payment environment.

The predictions and recommendations produced by this project should therefore be treated as analytical decision-support outputs rather than guaranteed fraud determinations.