Demand Model Generation from Traces: Adaptive KDE Data-Driven Optimization

Nowadays IoT and big data technologies make it easy to collect a huge amount of precise data about geospatial mobility. This abundant data improves mobility systems' study and optimization by allowing the application of data-driven approaches to gather accurate mobility models. In this paper, we evaluate techniques based on the optimization of Kernel Density Estimation (KDE) for modelling and generalizing the observed spatial data. We show how traditional KDE models fail to generalize the information obtained from mobility datasets, if not properly tuned. This mostly comes from the heterogeneity of the data, which changes over time (e.g., day and night) and space (e.g., dense or suburban areas). These issues call for fine grained parameter tuning, which we solve by automatically finding the optimal parameters for classical and variable KDE (V-KDE) approaches. We extensively compare the models' errors in capturing the mobility patterns using car sharing data. Our results show how optimized KDE and V-KDE models can better capture mobility patterns than traditional KDE approaches, both in 2D (i.e., origin or destination only) or 4D (i.e., representing origin-destination matrices) scenarios. We offer our tool available as open-source as a python module.