Predicting the technological complexity of global cities based on unsupervised and supervised machine learning methods

Analyzing and predicting innovation in global cities, i.e. cities with a high degree of economic integration into the world economy, can help identify emerging technologies and inform investment decisions that facilitate talent attraction and urban planning. In this context, the contribution of this paper is to analyze the technological complexity of global cities. We show how the combination of state-of-the-art network community detection and supervised machine learning can support local innovation and development policies by predicting the future competitiveness of global cities based on an up-to-date patent dataset. Network community detection with the Poisson stochastic block model is used as an unsupervised pre-processing step to find cities with similar innovation profiles and create homogeneous training sets that improve predictive power, interpretability and computational efficiency in a subsequent supervised learning task. The paper then compares the use of different supervised machine learning methods to predict the future competitiveness of global cities. Tree-based methods turn out to achieve better prediction performance than other supervised machine learning methods on various metrics based on the ground truth derived from historical patent production. The analytical method used in this paper can help policy makers identify technology sectors where global cities could focus their future investments and provide information on the temporal evolution of geographical patterns related to innovation.