The extended abstract: Parametric sensitivity in Graph Neural Network for Urban Cluster Detection, a case study has been accepted in the 10th International Conference on Complex Networks and their Applications.
Urban cluster detection is a relevant task in urban data analysis, with applications in urban policy. Many urban clustering methods have been proposed, but it is unclear whether the clusters inferred by the different methods are stable or dependent on an adequate selection of hyperparameters.
In this paper, we use a graph neural network-based clustering (GNNC) approach to uncover urban structure using geocoded data on socioeconomic status (SES) and aesthetic characteristics of the urban environment. We then perform a case study using data from Santiago, Chile, to examine the stability of the clusters inferred by the GNNC method with respect to changes in the data aggregation scheme and the model’s hyperparameters. Santiago is a city with high segregation, where most high-income households are located in the North-East, making it a suitable place to study parametric sensitivity. Results shows that clusters in the study case do not suffer from substantial variations for different parameter specifications.
You can find more detail in the Complex Nexworks 2021 website.