关键词: 流动人口； 社会融入； 市民化； 机器学习

Abstract: Against the background of China's people-centered new-type urbanization, a central policy challenge is how to promote the long-term social incorporation of internal migrants. This issue is especially salient where access to welfare, housing, and local public services remains uneven. Existing studies have identified institutional, social, and individual determinants of migrant incorporation, but have focused mainly on integration levels, linear relationships, and group comparisons. Less is known about how key factors operate through non-linear effects, threshold changes, and interaction mechanisms, and whether different dimensions of integration follow different pathways. To address these questions, this study examines the dynamic mechanisms, non-linear logic, and systemic characteristics of migrant social integration in China.

The analysis uses data from the resident questionnaires of the “Jiexiang Zhongguo” Urban Survey project organized by Renmin University of China in 2023 and 2024. After screening resident-level questionnaires, the study focuses on internal migrants who reside in destination communities but do not hold local hukou, yielding 1839 observations from 31 provincial-level units and 300 cities. The study employs CatBoost to model non-linear relationships and handle missing values, and applies SHAP values to interpret variable importance, marginal effects, and interaction patterns. Recursive feature elimination with cross-validation retains 21 explanatory variables. The benchmark CatBoost model reports an F1 score of 0.75513, an accuracy of 0.75989, and a recall of 0.76766.

The results show that migrant social integration is a composite process shaped jointly by institutional support, residential conditions, spatial mobility, time allocation, and digital participation. Community service satisfaction, insurance coverage, housing property rights, internet use frequency, and living arrangements constitute the main drivers, whereas economic gradient, migration distance, working hours, and destination-city hierarchy constitute the principal constraints. The nine most important variables account for 59.14% of the total contribution. These factors display threshold effects, non-linear marginal changes, and interaction patterns characterized by buffering, substitution, and reinforcement. Interaction analysis indicates that internet use, community services, and insurance coverage can mitigate disadvantages associated with migration distance and economic gradient; that positive conditions may substitute for one another when one support is already sufficient; and that long-distance mobility and long working hours can become more restrictive under certain conditions. Further analysis indicates that structural integration places greater weight on institutional acquisition, organizational participation, and social embeddedness, whereas cultural integration is more closely associated with identity, belonging, subjective acceptance, community service satisfaction, residential stability, and everyday living environments. This distinction shows that “entering urban society” and “identifying with urban society” are not identical processes.

Rather than treating social integration as a single, linearly determined outcome, the analysis specifies its internal differentiation and traces how multiple factors interact under certain conditions. At the theoretical level, it clarifies the internal structure of migrant social integration and distinguishes structural incorporation from cultural incorporation. At the practical level, the findings indicate that promoting fuller migrant incorporation requires coordinated community-based public services, portable social insurance, housing support, digital inclusion, and labor protection, while expanding neighborhood interaction and community participation so that migrants can move from “entering urban life” toward “identifying with urban society.”

Keywords: Domestic Migrants, Social Integration, Citizenization, Machine Learning