Toward Understanding Hidden Patterns in Human Mobility Using Wi-Fi

Abstract

A reliable model for identifying spatial-temporal regularities in human dynamics is rewarding in many applications such as computer networking and mobile communication. These hidden patterns are inherited from our repeating behaviours with respect to three primary contexts: time, space, and social environments. Thus, selecting a suitable source of sensor data that is scalable, multidimensional, and social network illustrative, can enable us to develop a reliable human mobility model and potentially a prediction system. We first demonstrate that collected Wi-Fi network scans from mobile phone devices share a similar set of characteristics to real-world large-scale networks. One aspect particularly is the long-tailed property of node degree distribution of projection networks. This feature can be interpreted as the robustness of the system against structural changes caused by removing a set of nodes or connections. Then, we transform Wi-Fi events into a tabular data format containing different time granularities and location-tagged information. However, the new data is sparse and difficult to analyze. Thus, we reduce the dimensionality of the data by extracting its structural patterns using the principal components of the new features. Our analysis shows that we can reconstruct the original data with more than 90% accuracy using only a set of top eigenvectors with one-quarter of the original features, while the outliers or the noisy user data are filtered out. Our proposed technique helps to visualize user similarities and behaviour dynamics, and reduce the computation complexity of further analysis.