Discovering user context with mobile devices : location and time

Abstract

Life for many people is based on a set of daily routines, such as home, work, and leisure. If the activities in life occur in recurring patterns, then the context in which they occur should also follow a pattern. In this thesis, we explore using cell phones for learning recurring locations using only a timestamped history of the cell tower the device is connected to. We base our approach on an existing graph-based online algorithm, but modify it to compute additional statistics for offline analysis to obtain better results. We then further refine the offline algorithm to include time-partitioned nodes to resolve some observed shortcomings. Finally, we evaluate all three algorithms on a dataset of GSM readings over a one month period, and show how our successive modifications improved the locations found.