Predicting neighborhoods’ socioeconomic attributes using restaurant data

Abstract

Accessing high-resolution, timely socioeconomic data such as data on population, employment, and enterprise activity at the neighborhood level is critical for social scientists and policy makers to design and implement location-based policies. However, in many developing countries or cities, reliable local-scale socioeconomic data remain scarce. Here, we show an easily accessible and timely updated location attribute—restaurant—can be used to accurately predict a range of socioeconomic attributes of urban neighborhoods. We merge restaurant data from an online platform with 3 microdatasets for 9 Chinese cities. Using features extracted from restaurants, we train machine-learning models to estimate daytime and nighttime population, number of firms, and consumption level at various spatial resolutions. The trained model can explain 90 to 95% of the variation of those attributes across neighborhoods in the test dataset. We analyze the tradeoff between accuracy, spatial resolution, and number of training samples, as well as the heterogeneity of the predicted results across different spatial locations, demographics, and firm industries. Finally, we demonstrate the cross-city generality of this method by training the model in one city and then applying it directly to other cities. The transferability of this restaurant model can help bridge data gaps between cities, allowing all cities to enjoy big data and algorithm dividends.

Economic outcomes predicted by diversity in cities

“…This dataset is all publicly available and downloaded from Meituan-Dianping website. Dong et al [31] use the same Dianping dataset to accurately predict socioeconomic attributes of various neighborhoods in several Chinese cities including Beijing. The attributes are daytime and nighttime population, number of firms, and consumption level in those areas.…”

Estimation of Socioeconomic Attributes from Location Information

“…Before defining ROC AUC and PR AUC, we introduce the false positive rate, r fp , and true positive rate, r tp , r fp = FP TN+FP (9) r tp = TP TP+FN ,…”

Estimation of Regional Economic Development Indicator from Transportation Network Analytics

“…The relationship between the fitted city GDP values using the multiple linear regression (MLR) models of transportation features and the actual city GDP in three provinces (i.e., Liaoning, Jiangsu, and Shaanxi) are summarized in Fig , S4, and S5, the residuals center on zero and are not correlated with any predictors, which indicate that these models' predictions have a relatively constant variance and show homoscedasticity and normality. In addition, by applying two regularized regression methods: the Ridge regression 46 and the least absolute shrinkage and selection operator (LASSO) regression 47,48 , the smallest RMSE values were 54.2 (Liaoning), 121.0 (Jiangsu), and 30.66 (Shaanxi) billion CNY, respectively. Moreover, these two methods only select a few predictors while reducing the coefficients of other highly correlated predictors to zero.…”

“…By shrinking some coefficients, the Ridge regression and the LASSO regression are able to control the multicollinearity in the model. It tends to pick one predictor from a few very correlated predictors and set the coefficients of the others to zero 47,48 . Therefore, the regularization techniques are very helpful when there are many intercorrelated features and feature selection is necessary 64 .…”