Policy learning, as a deep integration of artificial intelligence technologies with causal inference theory, provides a scientific foundation for precision social governance in the era of big data. Grounded in the practical context of China's large-scale governance and complex policy constraints, this paper systematically reviews the international frontier literature on policy learning and proposes a policy learning algorithm based on convexification theory. This approach effectively addresses the computational optimization challenges inherent in complex, high-dimensional policy spaces, thereby significantly improving policy learning efficiency under big data and multiple-constraint settings. The paper further conducts an empirical analysis using China's Minimum Livelihood Guarantee (Dibao) program as a case study, drawing on data from the China Household Finance Survey (CHFS). The results demonstrate that the proposed method can generate highly interpretable and precisely targeted Dibao allocation schemes, effectively expanding social welfare gains and enhancing governance efficiency. This study provides an important reference for leveraging cutting-edge digital and intelligent technologies to advance the modernization of social governance in China.