Combining space syntax with machine learning to predict seating places: the case of Gulbenkian estate in Portugal

A recent trend in space syntax is the application of machine learning (ML) techniques to extend the analyses traditionally performed with methods from graph theory. The goal of this paper is to explore the relationship between visual syntactic measures and the location of seating places in a grid with the resort of ML methods. As far as a public housing estate (Bairro Gulbenkian) located in Odivelas, near Lisbon, Portugal, is concerned, we found that the location of benches can be predicted accurately from visual connectivity, clustering, control, controllability, integration and through vision using partial least squares or random forests. In fact, these two methods provide a better balance between sensitivity (the proportion of seating places classified as such) and specificity (the proportion of other places classified as such) than logistic regression, least absolute shrinkage operator (LASSO), decision trees, support vector machines and neural networks. In addition, we found that visual clustering, integration, control and through vision may be the key measures to predict seating places.