Distributed Human Trajectory Sensing and Partial Similarity Queries

Advances in wireless communication technology have allowed for the collection of large-scale human motion trajectories by recording the appearance of mobile devices within the neighborhood of wireless base stations. Such city-scale datasets pose new challenges on efficient data collection, analysis and similarity based queries. In this paper, we propose new partial similarity measures, categorized as time-sensitive, order-sensitive and order-insensitive ones, and show with real data that these partial similarity measures are more robust than classical measures and more suitable for generating meaningful query results in near-neighbor type of data mining applications. Further, the power of the partial similarity persists even with significant down-sampling. We presented rigorous analysis of the performance of partial similarity measures with subsampling. Our evaluation using real data shows high recall and precision (≥ 90%) with samples only in the order of 1% of the original data size.