Class-Imbalanced Poverty Recidivism Prediction under Rural Revitalization Contexts

Under the background of the consolidation of poverty alleviation achievements, preventing large-scale poverty relapse has emerged as a critical challenge. To address class imbalance problem in poverty recidivism prediction, this study proposes a cost-sensitive LightGBM (cLightGBM) based poverty recidivism prediction model. First, the SMOTENC oversampling algorithm is employed to reconstruct sample space distribution, resolving class imbalance in mixed-feature spaces. Subsequently, a cost-sensitive LightGBM model is designed to enhance the model’s discriminative capability for high-risk samples. Experimental validation utilizes a monitoring dataset of poverty-alleviated households from a digital government platform, containing 2,327 samples with 36 feature variables across four dimensions: personal attributes, household economic status, village development level, and government assistance measures. Results demonstrate that the cLightGBM model achieves significant performance improvement with accuracy increased from 93.56% to 97.613% and F1-score from 68.299% to 97.613%. Interpretability analysis reveals that labor force size, distance to village main roads, and housing area constitute critical variables for poverty recidivism prediction.