Machine learning-based prediction of family return to poverty in the critical poverty groups

To consolidate the achievements of poverty alleviation, identifying the likelihood of return to poverty among the critical poverty groups is essential for proactive intervention. This study aims to predict the return to poverty of family in these groups, based on personal, family, and community environmental characteristics. Employing XGBoost and SHAP decomposition on data collected from the China Family Panel Studies between 2016 and 2020, we find that a combination of individual, family, and community environmental characteristics can effectively predict the likelihood of families in the critical poverty groups returning to poverty. The predictive performance of the family characteristics model is better than that of the personal and community environmental characteristics models. Among families in the critical poverty groups, those with higher expenditures on food, culture, education, and recreation are less likely to return to poverty, while those with lower satisfaction with the surrounding environment are prone to such an outcome. Additionally, most characteristics exhibit a nonlinear relationship with the return to poverty. This study broadens the understanding of family return to poverty from a predictive perspective, and provides empirical evidence to inform proactive supervision and policy intervention. First published online 13 November 2026