Automated Detection of Trajectory Groups Based on SNN-Clustering and Relevant Frequent Itemsets

Classification has been proposed for the automated detection of similarity groups in spatio-temporal data. However, recent approaches have introduced clustering based solutions to avoid the huge overhead for the manual classification of training and test data. This paper presents a combination of shared nearest-neighbor clustering and an adapted search for frequent itemsets to not only find similarity groups in sets of trajectories called team moves but also clusters of similar individual trajectories.Dynamic Time Warping is introduced as the underlying notion of trajectory distance on which the notion of trajectory similarity will be defined. Since the search for frequent itemsets is used to find similarity groups of team moves, an explicit distance criterion for team moves can be avoided. However, a notion of relevance is introduced that allows to distinguish trajectories with an impact on team moves from others. In addition, the paper will introduce enhanced quality indexes for shared nearest neighbor based trajectory clustering that allow to compare parameter settings in order to find the optimal clustering solution for a given problem.