Modeling Socioeconomic Variability Using Geographic Factors: A Comparative Study of Network-Based Model Architectures in Monocentric and Polycentric Metropolitan Regions

Abstract

With 70 percent of the global population expected to live in urban regions by 2050 [53], it is imperative to better understand how the geography of these urban spaces relates to the socioeconomic characteristics of their occupants. This paper presents a set of computational models implementing linear, multi-layer perceptron, and graph convolutional neural network architectures that utilize census tract geographic data in the New York City and Los Angeles metropolitan areas to predict census tract socioeconomic data. While the high accuracy of all linear, multi-layer perceptron, and graph convolutional neural network models provides convincing evidence of the role of geographic factors in predicting socioeconomic variability in urban spaces, small differences between monocentric, New York City, and polycentric, Los Angeles, metropolitan regions emerged. The multi-layer perceptron models provide the highest accuracy in predicting socioeconomic variability in both monocentric and polycentric urban spaces. However, the graph convolutional neural network models trained on monocentric metropolitan area data demonstrate the most flexibility by maintaining high accuracy when predicting socioeconomic data in unfamiliar polycentric urban spaces.