MFPMiner: Mining Meaningful Frequent Patterns from Spatio-textual Trajectories

Abstract

In the second decade of this century, technical progress has led to a worldwide proliferation of devices for tracking the movement behavior of a person, a vehicle, or another kind of entity. One of the consequences of this development is a massive and still growing amount of movement and movement-related data recorded by cellphones, automobiles, vessels, aircraft, and further GPS-enabled entities. As a result, the requirements for managing and analyzing movement records also increase, serving commercial, administrative, or private purposes. Since the development of hardware components cannot keep pace with the data growth, exploring methods of analyzing such trajectory datasets has become a very active and influential research field. For many application scenarios, besides the spatial trajectory of an entity, it is desirable to take additional semantic information into consideration. These descriptions also change with time and may represent, e.g., the course of streets passed by a bus, the sequence of region names traversed by an aircraft, or the points of interest in proximity of the positions of a taxi. Such data may be directly recorded by a sensor (such as the altitude of an aircraft) or computed from the spatial trajectory combined with some underlying information (for example, street names). It is often helpful or even necessary to focus on such semantic information for efficient analyses, as changes usually occur less frequently than it is the case for the spatial trajectory, where data points usually arrive in very close temporal distances. However, any kind of querying requires a deep semantic knowledge of the dataset at hand, particularly for retrieving the set of trajectories that match a certain mobility pattern, that is, a sequence of temporal, spatial, and semantic specifications. In this article, we introduce a framework named MFPMiner 1 for retrieving all mobility patterns fulfilling a user-specified frequency threshold from a spatio-textual trajectory dataset. The resulting patterns and their relative frequency can be regarded as a knowledge base of the considered data. They may be directly visualized or applied for a pattern matching query yielding the set of matching trajectories. We demonstrate the functionality of our approach in an application scenario and provide an experimental evaluation of its performance on real and synthetic datasets by comparing it to three competitive methods. The framework has been fully implemented in a DBMS environment and is freely available open source software.