Trajectory Feature Stores for Privacy-Preserving Mobility Intelligence: Schema Design, Quality Control, and Federated Benchmarking

Abstract

Urban mobility intelligence increasingly depends on high-frequency GPS traces, app-based location events, and fleet trajectories. Yet the same signals that support traffic forecasting, destination prediction, and transport-service optimization also expose sensitive behavioral routines. This article develops a trajectory feature store framework for privacy-preserving mobility intelligence. The framework converts raw trajectory events into governed, versioned, and quality-controlled feature views that can be used in federated learning without centralizing raw location histories. It specifies a schema design for spatial, temporal, behavioral, and privacy metadata; a quality-control layer for completeness, spatial validity, temporal continuity, label reliability, and client balance; and a federated benchmarking layer for comparing model utility, communication overhead, and residual privacy risk under controlled feature snapshots. A synthetic analytical evaluation inspired by GeoLife and T-Drive-style mobility settings shows that feature-store controls improve average feature quality from 0.77 to 0.90, reduce training-serving inconsistency, and support more stable privacy-utility trade-offs under differential privacy and secure aggregation. The study contributes a data-governance-centered perspective on federated trajectory mining by shifting attention from model architecture alone to the reproducibility, auditability, and operational readiness of the mobility feature layer.

Keywords: trajectory feature store; federated learning; mobility intelligence; data governance; differential privacy; quality control