Unsupervised context detection using wireless signals

Abstract

The sensing context plays an important role in many pervasive and mobile computingapplications. Continuing from previous work [D. Phung, B. Adams, S. Venkatesh,Computable social patterns from sparse sensor data, in: Proceedings of First InternationalWorkshop on Location Web, World Wide Web Conference (WWW), New York, NY,USA, 2008, ACM 6972.], we present an unsupervised framework for extracting usercontext in indoor environments with existing wireless infrastructures. Our novel approachcasts context detection into an incremental, unsupervised clustering setting. Using WiFiobservations consisting of access point identification and signal strengths freely availablein office or public spaces, we adapt a density-based clustering technique to recover basicforms of user contexts that include user motion state and significant places the user visitsfrom time to time. High-level user context, termed rhythms, comprising sequences ofsignificant places are derived from the above low-level context by employing probabilisticclustering techniques, latent Dirichlet allocation and its n-gram temporal extension. Theseuser contexts can enable a wide range of context-ware application services. Experimentalresults with real data in comparison with existing methods are presented to validate theproposed approach. Our motion classification algorithm operates in real-time, and achievesa 10% improvement over an existing method; significant locations are detected withover 90% accuracy and near perfect cluster purity. Richer indoor context and meaningfulrhythms, such as typical daily routines or meeting patterns, are also inferred automaticallyfrom collected raw WiFi signals.