Modeling Visit Potential of Geographic Locations Based on Mobility Data

Every day people interact with the environment by passing or visiting geographic locations. Information about such entity-location interactions can be used in a number of applications and its value has been recognized by companies and public institutions. However, although the necessary tracking technologies such as GPS, GSM or RFID have long found their way into everyday life, the practical usage of visit information is still limited. Besides economic and ethical reasons for the restricted usage of entity-location interactions there are also two very basic problems. First, no formal definition of entity-location interaction quantities exists. Second, at the current state of technology, no tracking technology guarantees complete observations, and the treatment of missing data in mobility applications has been neglected in trajectory data mining so far. This thesis therefore focuses on the definition and estimation of quantities about the visiting behavior between mobile entities and geographic locations from incomplete mobility data. In a first step we provide an application-independent language to evaluate entity-location interactions. Based on a uniform notation, we define a family of quantities called visit potential, which contains the most basic interaction quantities and can be extended on need. By identifying the common background of all quantities we are able to analyze relationships between different quantities and to infer consistency requirements between related parameterizations of the quantities. We demonstrate the general applicability of visit potential using two real-world applications for which we give a precise definition of the employed entity-location interaction quantities in terms of visit potential. Second, this thesis provides the first systematic analysis of methods for the handling of missing data in mobility mining. We select a set of promising methods that take different approaches to handling missing data and test their robustness with respect to different scenarios. Our analyses consider different mechanisms and intensities of missing data under artificial censoring as well as varying visit intensities. We hereby analyze not only the applicability of the selected methods but also provide a systematic approach for parameterization and testing that can also be applied to the analysis of other mobility data sets. Our experiments show that only two of the tested methods supply unbiased estimates of visit potential quantities and are applicable to the domain. In addition, both methods supply unbiased estimates only of a single quantity. Therefore, it will be a future challenge to design methods for the entire collection of visit potential quantities. The topic of this thesis is motivated by applied research at the Fraunhofer Institute for Intelligent Analysis and Information Systems IAIS for business applications in outdoor advertisement. We will use the outdoor advertisement scenario throughout this thesis for demonstration and experimentation.