Mobility data disaggregation: a transfer learning approach

In this paper we present a machine learning approach to derive temporal disaggregated origin-destination matrices from static censuses on population mobility data. A better understanding of the displacements over a land is an issue for decision making and territorial planning. The primary objective of this paper is to transform census data towards a disaggregated form, allowing its merging with other mobility data. Most of Human displacements are linked to Home and Work locations, and these commuting patterns regularly generate road congestion problems. Knowing the temporality of commuting patterns eases the decision making processes from local authorities. Mobility censuses often contain these Home and Work information, but are temporally aggregated. We thus propose a machine learning model that learns the temporal distribution of displacements from other mobility sources and allows to temporally disaggregate the displacements of such censuses. This model have been validated and showed its efficiency in a real context. The main advantage of our approach is that the resulting matrices inherit all the attributes contained in the aggregated census. Finally, the model is optimized so that it can be applied to other places wherever censuses on mobility data are available, and without additional computations.