Predicting Tax Defaults Through Feature Transformation and XGBoost Optimization

This study focuses on predicting tax defaults using advanced machine learning techniques, specifically Feature Transformation and XGBoost Optimization. Accurate prediction of tax defaults is crucial for improving tax collection efficiency and minimizing revenue losses. The analysis begins with data collection from the Kaggle Individual Income Tax Statistics dataset, which includes detailed taxpayer income, deductions, and credits. To enhance the predictive power of the model, Min-Max Normalization is applied to ensure all features are scaled uniformly, preventing larger values from dominating the model’s learning process. Following normalization, Principal Component Analysis (PCA) is utilized to reduce dimensionality by extracting the most significant features, which helps in simplifying the dataset while preserving essential information. XGBoost, a powerful gradient boosting algorithm, is then employed for predicting tax defaults. XGBoost’s strength lies in its ability to handle complex relationships between features and mitigate issues like overfitting and data imbalance, which are common in tax default datasets where the majority of taxpayers do not default. The model is fine-tuned through hyperparameter optimization to further improve prediction accuracy. This study demonstrates the effectiveness of combining feature transformation techniques with XGBoost to create a robust and scalable solution for tax default prediction. Feature importance analysis is also conducted to identify key drivers of tax defaults, providing valuable insights for tax authorities to implement preemptive measures. The proposed method is implemented in Python and has an accuracy of about 98.96% which is superior than BiLSTM, LSTM and CNN.