Public mobility sharing systems are an important component of sustainable transport, particularly for last-mile travel. However, analysing trip patterns using open standards such as GBFS can be challenging due to vehicles frequently being assigned new identifiers and missing GPS trajectories, preventing a detailed tracking. To overcome this limitation, we present a machine learning pipeline that retrospectively predicts trip destinations within this circumstances—making it possible to partially recover travel patterns for GBFS data.
Our approach involves a three-step prediction pipeline: (1) candidate generation and reduction using spatial-temporal filtering; (2) multi-target regression via XGBoost to estimate destination coordinates; and (3) selection of the best-matching candidate. Our approach achieves an average accuracy of 77% across five German and 74% across five international cities within a tolerance of 500 metres. Compared to existing approaches, our method improves prediction accuracy by an average of 20% over methods that also do not use user-specific or GPS trajectory features.
These results demonstrate the feasibility of accurately predicting destinations in shared mobility despite rotating vehicle identifiers and missing trajectory data, thereby supporting improved system analysis and planning.
Dieser Eintrag ist Teil der Universitätsbibliographie.
Das Dokument wird vom Publikationsserver der Universitätsbibliothek Mannheim bereitgestellt.
Creative Commons Namensnennung 4.0 International (CC BY 4.0)
Suche Autoren in
