Bidirectional imputation of spatial GPS trajectories with missingness using sparse online Gaussian Process.

Area of Science:

• Mobility Data Analysis
• Geospatial Information Systems
• Computational Epidemiology

Background:

• Smartphone GPS data is often sparse (<10% sampling) to conserve battery, leading to significant missing mobility information.
• Existing trajectory imputation methods face high computational complexity and may not accurately represent real-world movement.

Purpose of the Study:

• To develop an efficient and accurate method for imputing missing GPS trajectory data.
• To address the computational challenges of existing methods while preserving data accuracy.

Main Methods:

• A novel bidirectional Gaussian Process imputation algorithm utilizing sparse online Gaussian Process.
• The method accounts for the spherical geometry of Earth and incorporates condition checks for enhanced accuracy.
• Linear time complexity relative to sample size for efficient online processing.

Main Results:

• Imputed trajectories closely mimic real-world mobility patterns.
• Accurate estimation of daily/hourly summary statistics (e.g., time spent at home, distance traveled) with confidence intervals covering ground truth.
• Demonstrated significant speed improvements over existing methods for long-term observations (>3 months).

Conclusions:

• The proposed method outperforms existing approaches in both accuracy and speed.
• Enables continuous analysis of mobility data for applications like behavioral anomaly detection and epidemic contact tracing.
• Provides guidelines for optimizing GPS sampling strategies.