Cross-domain transfer in mobility-based socio-demographic prediction: Evidence from French and Swiss mobility datasets

Abstract

The widespread adoption of geolocation technologies has transformed the analysis of human mobility, offering a scalable alternative to traditional household travel surveys (HTS). However, such datasets typically lack socio-demographic labels, making it challenging to assess their representativeness. In this paper, we address this issue through an indirect evaluation strategy based on cross-domain model transfer. Models trained to infer socio-demographic attributes from mobility trajectories are applied to an unlabeled target dataset, where they are used to reconstruct aggregate population distributions. Using French and Swiss mobility datasets, we evaluated the robustness of this transfer approach. Our results show that dataset shifts significantly limit model portability, leading to substantial degradation in individual-level predictive performance. We then investigate quantification as a post hoc correction for the prior shift and demonstrate that Probabilistic Classify and Count (PCC) remains effective in estimating class prevalence under distributional changes. Finally, we highlight promising directions for improving cross-domain generalization, including the learning of domain-invariant representations and the development of alignment strategies to address covariate and concept shifts.