Abstract
Urban research has long recognized that neighbourhoods are dynamic andrelational. However, lack of data, methodologies, and computer processing powerhave hampered a formal quantitative examination of neighbourhood relationaldynamics. To make progress on this issue, this study proposes a graph neuralnetwork (GNN) approach that permits combining and evaluating multiple sourcesof information about internal characteristics of neighbourhoods, their pastcharacteristics, and flows of groups among them, potentially providing greaterexpressive power in predictive models. By exploring a public large-scaledataset from Yelp, we show the potential of our approach for consideringstructural connectedness in predicting neighbourhood attributes, specificallyto predict local culture. Results are promising from a substantive andmethodologically point of view. Substantively, we find that either local areainformation (e.g. area demographics) or group profiles (tastes of Yelpreviewers) give the best results in predicting local culture, and they arenearly equivalent in all studied cases. Methodologically, exploring groupprofiles could be a helpful alternative where finding local information forspecific areas is challenging, since they can be extracted automatically frommany forms of online data. Thus, our approach could empower researchers andpolicy-makers to use a range of data sources when other local area informationis lacking.