Explaining a century of Swiss regional development by deep learning and SHAP values

Abstract

We use a graph convolutional neural network (GCN) for regional development prediction with population, railway network density, and road network density of each municipality as development indicators. By structuring the long-term time series data from 2833 municipalities in Switzerland during the years 1910–2000 as graphs over time, the GCN model interprets the indicators as node features and produces an acceptable prediction accuracy on their future values. Moreover, SHapley Additive exPlanations (SHAPs) are used to make the results of this approach explainable. We develop an algorithm to obtain SHAP values for the GCN and a sensitivity indicator to quantify the marginal contributions of the node features. This explainable GCN with SHAP decomposes the indicator into the contribution by the previous status of the municipality itself and the influence from other municipalities. We show that this provides valuable insights into understanding the history of regional development. Specifically, the results demonstrate that the impacts of geographical and economic constraints and urban sprawl on regional development vary significantly between municipalities and that the constraints are more important in the early 20th century. The model is able to include more information and can be applied to other regions and countries.