Classifying Urban Fabrics into Mobile Call Activity with Supervised Machine Learning

Strong spatial relationships exist between call volume and human activities. In this paper, we show machine learning models can classify base station coverage areas into classes of mobile call profile based on areas geographic and demographic properties. We propose the novel approach of taking clustered Call Detail Records (CDR) based time series as ground truth, and passing heterogeneous features as inputs. These features were land use and points of interest retrieved from OpenStreetMap database, and demographic data provided Facebook Research. Together, they allowed the creation of a representation of urban fabric. CDR clusters were then characterized by these features through SHAP model interpretation. Three models were tested in the region of Dakar with CDR coming from D4D-Senegal Challenge. Results exceeded the accuracy of systematic most common class classification and the accuracy of models trained only on population dataset. Additionally, models generalization capacity were evaluated in Thies, with results equalling those of the baseline.