Prediction and pattern recognition of the large-scale poverty-returning risks: empirical evidence from machine learning

Abstract How to effectively predict the risk of large-scale poverty-returning and realize prior targeted intervention are essential measures to prevent large-scale poverty-returning. In view of this, based on the Sustainable Livelihoods Framework, this paper employs machine learning algorithms to study the prediction and pattern recognition of large-scale poverty-returning risks using the China Family Panel Studies (CFPS) and satellite remote sensing data. Our findings demonstrate a remarkable prediction accuracy rate of 91.23% for large-scale poverty-returning, employing the random forest algorithm. Notably, key contributing variables to this prediction encompass education, health, and government subsidies, highlighting the potential of supervised learning methods in predicting large-scale poverty-returning risk. Furthermore, the K-means clustering analysis reveals three discernible risk patterns within large-scale poverty-returning, including low-risk, three-risk, and five-risk patterns. The findings could provide scientific empirical evidence for the governance of different types of large-scale poverty-returning risks.