Community-level education percentile rank estimation in China using multi-source big data and machine learning

Socio-economic data with fine-grained spatial resolution forms the basis of socio-spatial analysis and policymaking. In response to the limited availability of such data in China, this study provides an open-access, community-level dataset on education percentile rank - a more accurate indicator of social status than years of education. Our dataset comprises 122,126 communities, covering 97.9% of prefecture-level administrative units and 81.8% of county-level administrative units. The data is estimated using an XGBoost machine learning model based on the relationship between mean education percentile rank and the characteristics of the built environment, including functions and facilities, street scene elements, vitality, human perception, physical disorder, and topography at the community level. Multi-source data, including the Chinese General Social Survey, points of interest, road networks, night-time lighting, and street view images processed using computer vision techniques such as semantic segmentation, object detection, and image regression, are used for model training and inference. Our final education predictions are highly accurate at prefecture, county, and community levels. This dataset enables fine-grained socio-spatial analyses across disciplines.