Contributions to the understanding of human mobility and its impact on the improvement of lightweight mobility prediction algorithms

Resumo

Human mobility is key in fields like urban planning, protocols for mobile networks, or service personalization, among others. Besides, a large number of studies emerged in the last years thank to more complete mobility data sets, coming from the use of mobile phones as mobility proxies that continuously record their owners' locations. Traditionally, many of the applications of human mobility, like service personalization, were based on the current location of the user. However, this focus has recently started to shift to the user's mobility habits and her future location. This change allows having time in advance to provide services related to the usual habits of the user. This thesis focuses on broadening the understanding of human mobility through the analysis of the location data recorded by mobile devices, and finding ways to increment the probability of making right predictions about their future locations. To confront this challenge, it is divided into three stages: the mobility data collection, the extraction and analysis of the mobility features reflected into the recorded data, and the analysis of a set of prediction algorithms to propose some improvements. The intrinsic privacy risks associated to the disclosure of the location and mobility data of the user are also considered. In the first stage, the analysis of the sensors available in mobile devices and the requirements of the thesis lead to choose the cellular network as the source of mobility data. After analyzing the existing data sets containing this kind of data, it is decided to carry out a new mobility data collection campaign to obtain a more complete data set. The second stage is focused on extracting mobility features from the data chosen in the previous step, and spot the biases introduced by the data collection scheme. In order to eliminate these biases, several filtering techniques are proposed to delete the maximum number of events not representing the movement of the user. For the next stage, the specific family of LZ-based prediction algorithms is chosen to analyze their results when using mobility data obtained using different schemes and then filtered. By leveraging the mobility features studied in the previous stage, and based on their relationship with the prediction results, several modifications of the original algorithms are proposed to increase the fraction of right predictions. Finally, in the privacy preservation plane, the shift from disclosing static location profiles to mobility profiles leads to the proposal of a new privacy metric, based on the concept of entropy rate. The goal is to consider both the spatial as well as temporal information in a mobility profile. Some privacy-enhancing perturbation techniques are tested with both location and mobility profiles using the new privacy metric, which unveils the noticeable amount of information stored in the temporal correlations.