Transforming urban fabric into mobile call traffic signatures
Résumé
Many studies have shown how to process mobile network data with machine learning algorithms to infer land uses or predict mobile traffic behavior. However, there are few, if any, machine learning applications to help deploy new cells. Current forecasting models are designed to predict the traffic of existing cells based on the historical data they produced. In a previous work, we have proposed a method to predict the class activity of a future cell, given its modeled area, demographic and geographic features. In this paper, we extend it by estimating static hour-by-hour median weekly activity aggregated at base station level, based on the same static inputs. We tested several machine learning models including transformers, compared their results and showed that our method beats simple baselines.