Contextual Location Imputation for Confined WiFi Trajectories

Abstract

The analysis of mobility patterns from large-scale spatio-temporal datasets is key to personalised location-based applications. Datasets capturing user location are, however, often incomplete due to temporary failures of sensors, deliberate interruptions or because of data privacy restrictions. Effective location imputation is thus a critical processing step enabling mobility pattern mining from sparse data. To date, most studies in this area have focused on coarse location prediction at city scale. In this paper we aim to infer the missing location information of individuals tracked within structured, mostly confined spaces such as a university campus or a mall. Many indoor tracking datasets may be collected by sensing user presence via WiFi sensing and consist of timestamped associations with the network’s access points (APs). Such coarse location information imposes unique challenges to the location imputation problem. We present a contextual model that combines the regularity of individuals’ visits to enable accurate imputation of missing locations in sparse indoor trajectories. This model also considers implicit social ties to capture similarities between individuals, applying Graph-regularized Nonnegative Matrix Factorization (GNMF) techniques. Our findings suggest that people’ movement in confined spaces is largely habitual and their social ties plays a role in their less frequently visited locations.