With the drastic increase of object trajectory data, the analysis and exploration of trajectories has become a major research focus with many applications. In particular, several approaches have been proposed in the context of similarity-based trajectory retrieval. While these approaches try to be comprehensive by considering the different properties of object trajectories at different degrees, the distance functions are always pre-defined and therefore do not support different views on what users consider (dis)similar trajectories in a particular domain.
In this paper, we introduce a novel approach to learning distance functions in support of similarity-based retrieval of multi-dimensional object trajectories. Our approach is more generic than existing approaches in that distance functions are determined based on constraints, which specify what object trajectory pairs the user considers similar or dissimilar. Thus, using a single approach, different distance functions can be determined for different users views. We present two learning techniques,
transformed Euclidean distance
transformed Dynamic Time Warping
. Both techniques determine a linear transformation of the attributes of multi-dimensional trajectories, based on the constraints specified by the user. We demonstrate the flexibility and efficiency of our approach with applications to clustering and classification on real and synthetic object trajectory datasets from different application domains.