CDR-based Trajectory Reconstruction of Mobile Network Data Using Transformers

With the development of telecommunication technologies, mobile devices, and data collected via mobile services, predicting individuals’ paths in cities has become of great interest. A common approach for achieving this goal is to build models based on sparse mobile network data recorded in Call Detail Records (CDRs) that can reconstruct the trajectories of individuals connected to the network. Recent techniques, such as the Switching Kalman Filter or Particle Filter, rely on probabilistic modeling. We introduce a novel deep-learning approach to this problem, drawing inspiration from the closely related sub-field of trajectory studies and the recent successful application of the Transformer architecture to the problem of map matching. In so doing, we create a synthetic CDR dataset based on real-life GPS trajectories of two individuals. Our framework involves training two Transformers sequentially -one used for map matching to obtain the road-level path from the underlying GPS traces, and the other tasked with reconstructing road segment sequences from the cell-level trajectory.