Discovering latent semantic structure in human mobility traces

European Conference on Wireless Sensor Networks |

Published by Springer

Human mobility is a complex pattern of movements and activities that are based on some underlying semantics of human behavior. In order to construct
accurate models of human mobility, this semantic behavior needs to be unearthed from the data sensed as a human being moves around and visits certain classes of locations such as home, work, mall, theater, restaurant etc. The ideal data for understanding the semantics of mobility would constitute timestamped mobility traces with detailed geographic locations with annotations about the type of each location. One way of achieving this is by following a hybrid strategy of participatory sensing (with each person carrying a wireless sensor device) and deploying static sensors at each location of interest – the contacts between the mobile and (annotated) static sensors can be logged at each location, and then collated to form an appropriate mobility traces. For example, a person can connect with his mobile phone over Bluetooth or WiFi to a local hotspot while checking into FourSquare at a restaurant. In the absence of static sensors, a person may manually annotate the places he visits on his device over time. However, most mobility traces consist of network connectivity data from cell phones (e.g., contact with towers) which lack detailed geographic locations and are ambiguous, noisy and unlabeled. Thus, it is important to extract the semantics of mobility that is latent in the available contact traces. To this end, we propose in this paper the concept of Probabilistic Latent Semantic Trajectories (PLST), an unsupervised approach to extract semantically different locations and sequential patterns of mobility from such traces. PLST extracts semantic locations as contextually co-occurring network elements (cell towers and Bluetooth devices) and models the behavior of their sequence. PLST extracts distinct locations with spatial, temporal and semantic coherency and can be used for accurate prediction of the next place a user visits. PLST also analyzes the complexity of mobility traces using information theoretic metrics to study the underlying structure and semantic content in mobility traces. This semantic content can be extracted allowing us to investigate mobility patterns in a completely unsupervised manner