An Extensive. Experimental Evaluation of Location Pre- diction Models For Moving participants (Chapter 6) As a part of our dynamic framework for task assignment, this chapter carries out an extensive experimental evaluation for existing mobility modeling approaches. Several location prediction approaches are proposed in literature for variety of location-based services and social networks, however, it is still challenging to adopt the most suitable algorithms for specific applications. While complex algorithms are developed every year, yet simple Markov Chain algorithms has shown promising results [67] in location prediction and remain as the main solution to many real-world applications. Existing location prediction approaches target different applications with their parameters tuned for certain data sets. Many of such approaches have not been studied in variety of experimental settings with different data, thus, they lack an extensive evaluation to show their strengths and limitations in different situations. Moreover, due to existing of variety of spatial and temporal modeling of trajectories, several approaches are not comparable directly [73], therefore, building a framework to evaluate and compare such methods using the same data and test parameters remains as a challenge. We perform an extensive evaluation of the existing state-of-the-art methods with the following contributions. • We identify different location prediction problems in literature and discuss their similarities and differences. We also give a formal definition of the general next location prediction problem which is studied widely with applications in real- world location-based services. • We categorize the existing approaches for next location prediction problem based on several aspects including the level of personalization, the spatial and temporal representation of trajectories, and the mobility behavior modeling scheme. • We study several state-of-the-art next location prediction algorithms and eval- uate them extensively using both synthetic and real-world data sets. We design a series of experiments to evaluate and compare these algorithms based on the criteria of accuracy, efficiency, and robustness using data with different features. • We use theoretical limits of human mobility prediction [88] to estimate the upper limits of predictability for our data sets. This feature allows us to identify the areas for improvement of algorithms in this study while comparing them to each other.
