Detecting Outlier Segments in Uncertain Personal Trajectory Data

Uncertainties such as missing data, noise, and outliers, are common in personal trajectory data due to extended indoor time. To improve the usability of such data, it is necessary to detect and repair errors. While smoothing-based methods can be used to deflate noise, such methods are not well suited to cleaning trajectory data in the presence of spatial outliers represented by abrupt change in location. We propose detecting spatial outliers at the level of segments using supervised machine learning algorithms. We partitioned a trajectory into segments based on time gaps (greater than recording interval) and used smoothed values of relevant features including movement state at the level of segments for detecting spatial outliers. Experimentation results indicate that Extreme Gradient Boosting outperforms Random Forest and Light Gradient Boosting Machine methods. The study indicates that gaps and movement state can be considered to partition a trajectory and improve outlier detection.