Precise Identification of Multi-Regional Relative Poverty: A Two-Stage Knowledge-Distilled Adaptive Framework

As poverty reduction strategies shift comprehensively toward alleviating relative poverty, precisely identifying multidimensional relative poverty populations has become critical for social governance. However, existing data-driven models often face algorithmic bottlenecks—such as spatial heterogeneity, regional sample sparsity, and extreme category imbalance—when applied to complex scenarios characterized by vast territories and significant regional disparities. To address this, this study proposes a two-stage knowledge-distilled adaptive gradient boosting framework (TKDAF). First, in the prior knowledge extraction stage (Stage I), a base structure-regularized gradient boosting tree model is constructed. Combined with SHAP game-theoretic attribution, this stage quantifies and extracts the global objective weights of poverty-inducing features across regions. Second, in the spatial adaptive enhancement stage (Stage II), the S-DAGB (Spatial-Adaptive Distilled Gradient Boosting) core prediction model is introduced. It achieves deep integration of multiple regularization and feature enhancement mechanisms by incorporating: feature space nonlinear reconstruction based on prior knowledge, a dynamic category weighting mechanism based on effective sample size (ENS), and spatial adaptive optimization using the TPE Bayesian algorithm.Empirical results based on the 2020 China Family Panel Studies (CFPS) multidimensional dataset demonstrate that the S-DAGB model not only effectively overcomes the generalization bottleneck of deep tree models in the sample-constrained Northeast region (achieving 93.52% accuracy),but also significantly improves precision in regions with highly heterogeneous features and extreme class imbalance, such as central and western China. This effectively reduces wasteful allocation of poverty alleviation resources caused by false positive errors. This study provides an algorithmic solution that combines high accuracy with interpretability for precise identification of relative poverty in complex data distribution scenarios.