Discovering human routines from cell phone data with topic models

Abstract

We present a framework to automatically discover people's routines from information extracted by cell phones. The framework is built from a probabilistic topic model learned on novel bag type representations of activity-related cues (location, proximity and their temporal variations over a day) of peoples' daily routines. Using real-life data from the Reality Mining dataset, covering 68 000+ hours of human activities, we can successfully discover location-driven (from cell tower connections) and proximity-driven (from Bluetooth information) routines in an unsupervised manner. The resulting topics meaningfully characterize some of the underlying co-occurrence structure of the activities in the dataset, including ldquogoing to work early/laterdquo, ldquobeing home all dayrdquo, ldquoworking constantlyrdquo, ldquoworking sporadicallyrdquo and ldquomeeting at lunch timerdquo.