A Learning-Based Approach for Multi-scenario Trajectory Similarity Search

Abstract

The ubiquity of positioning devices and wireless networks has been significantly boosting the development of LBS (location-based service) technology and applications. As is widely used in LBS, trajectory top-k query serves as the key operation in a variety of large-scale web services, such as route recommendation, user behavior pattern analysis, etc. Unfortunately, existing top-k methods usually measure the trajectory similarity from a certain perspective, which leads to niche applications, not to mention the high computation complexity causing performance bottlenecks. To enable more types of data mining and analysis for web services, we propose a learning-based approach for multi-scenario trajectory similarity search with the generic metric, named TrajGS. Different from existing trajectory similarity measurements that are calculated by a single principle, TrajGS aims at proposing a generic rule that considers the trajectory correlation from multiple aspects and improves the accuracy of search in any circumstances. Specifically, TrajGS has two innovative modules: 1) a new trajectory representation enabled by a bidirectional LSTM model that can better capture the contextual information of trajectory and overcome the position error caused by outliers. 2) A generic trajectory similarity metric combining multiple distance measures obtains more versatile and accurate top-k results in various scenarios. Extensive experiments conducted on real datasets show that TrajGS achieves both impressive accuracy and adequate training time on trajectory similarity search compared with single-metric solutions. In particular, it achieves 5x–10x speedup and 20%–30% accuracy improvement over Euclidean, Hausdorff, DTW, and EDR measurements.