Efficient and robust data augmentation for trajectory analytics: a similarity-based approach

Trajectories between the same origin and destination (OD) offer valuable information for us to better understand the diversity of moving behaviours and the intrinsic relationships between the moving objects and specific locations. However, due to the data sparsity issue, there are always insufficient trajectories to carry out mining algorithms, e.g., classification and clustering, to discover the intrinsic properties of OD mobility. In this work, we propose an efficient and robust trajectory augmentation approach to construct sizeable qualified trajectories with existing data to address the sparsity issue. The high-level idea is to concatenate existing trajectories to reconstruct a sufficient number of trajectories to represent the ones going across the OD pair directly. To achieve this goal, we first propose a transition graph to support efficient sub-trajectories concatenation to tackle the sparsity issue. In addition, we develop a novel similarity metric to measure the similarity between two set of trajectories so as to validate whether the reconstructed trajectory set can well represent the original traces. Empirical studies on a large real trajectory dataset show that our proposed solutions are efficient and robust.