Link prediction with local and global consistency preservation in spatio-temporal networks
University of New Brunswick
With the increasing deployment of connected positioning devices, we are witnessing the proliferation of connected data sets in the form of spatio-temporal networks such as Location-Based Social Networks (LBSNs), the Internet of Things (IoT), and smart transportation networks. Link prediction is a key research field in studying spatiotemporal networks as it improves our understanding of the underlying dynamics of the connected data sets by predicting missing or future links that represent the relations in a system. However, current research on link predictions in spatio-temporal networks has been mostly limited to friendship prediction in Location-Based Social Networks (LBSN), and even though local and global consistency have been regarded as important factors in predictive analytics, they have not yet been studied in spatio-temporal networks. One of the main research challenges is mainly related to addressing local consistency due to the substantial difference between the sense of locality in spatio-temporal networks in comparison to non-spatial networks. Moreover, incorporating the role of communities in link prediction in spatio-temporal networks specifically under the concepts of global consistency is another challenge that has not been addressed yet. These challenges have been addressed by proposing methods for carrying out link prediction with local and global consistency which are tested using data from two different shared-mobility systems namely bike-sharing and taxi systems from Chicago and New York City. Different prediction scenarios including the presence of periodic variations in the data and multi-step prediction have been considered. The comparison of the results from the proposed and baseline methods indicates that the proposed methods accurately predict the flow and other related variables (e.g., check-ins) in shared-mobility systems in different scenarios. For example, The proposed MFLOG model improves the bike-flow and check-in/out prediction error by 4.5% and 7.5% respectively, w.r.t baseline models. This can be associated with the successful design of the methods and consideration of local and global consistency in the model.