This study utilizes Gradient Boosting Regression Trees and Random Forest Agorithms to explore the predictive power of 64 features of six dimensions, which previous studies have identified as determinants of corporate tax avoidance. The results show that features of tax shelters have strong predictive power. Specifically, donation expenditure, total accruals, employment, and R&D expenditure emerge as the most influential predictors. However, external features have limited predictive power. Examining the relationships between key features and tax avoidance,this paper finds that, in general, donation expenditure and employment are negatively related to tax avoidance levels, while total accruals and R&D expenditure are positively associated with tax avoidance levels. However, in a few cases, employment is positively correlated with tax avoidance, and the correlations between tax avoidance and total accruals or R&D expenditure are weak. Interactive effectsexist between donation expenditure and return on total assets, total accruals and supervisory board size,employment and CEO tenure, and R&D expenditure and institutional ownership. Treebased machine learning algorithms have greater predictive power than linear regressions. Our results suggest that more attention should be paid to features of tax shelters in tax collection procedures and the design of intelligent tax systems, and that restricting tax shelters is crucial in policy making.