RoutineSense: A Mobile Sensing Framework for the Reconstruction of User Routines is the title of the Best Paper awarded at MobiQuitous 2015, the 12th EAI International Conference on Mobile and Ubiquitous Systems: Computing, Networking and Services, held on last July 22-24 at the Universidade de Coimbra.
Jean-Eudes Ranvier, Michele Catasta, Matteo Vasirani and Karl Aberer from the Ecole Polytechnique Fédérale de Lausanne (EPFL), presented their research about a system for the automatic reconstruction of complex daily routines from simple user states recorded by mobile applications.
Thanks to sensors such as GPS, accelerometer and microphone, several user-centric smartphone applications have been recently developed to collect data in order to understand our habits. This kind of applications focus on specific aspects of a user’s life, but what if we could identify a “routine”, from sequences of simple user states? This is what RoutineSense aims at.
The paper presents an incremental processing framework for fusing sensors data into simple states, which can be further aggregated into complex routines. The scope is to combine opportunistic sensing, unsupervised pattern generation and user feedback to discover and rank frequent and exceptional routines in an incremental way, following an established human memory mode.
Previous studies. Mobile sensing and mobile activity recognition have been studied in multiple context, so as the aggregation of simple activities into high-level routines. On the other hand, most of the previous studied methods are restricted to the detection of limited set of states. Moreover, other techniques uniquely based on GPS cannot benefit from additional semantic information and consider only sequential patterns, disregarding potential temporal swapping of events.
RoutineSense: how it works. The framework proposed by the EPFL team aggregates low level sensor readings into more abstract multimodal states which can be, in turn, aggregated into high level representation of the user routine. Furthermore, the algorithms used are essentially incremental, which allows for periodic processing of the increments as well as energy efficiency, both required for mobile applications. Inspired by the episodic memory model proposed by M. A. Conway, RoutineSense uses episodic elements (EE) which represent events or summary of events, in order to generate a high-level image of a user’s life based on his smartphone sensor readings. The EE can be aggregated into simple episodic memories (SEM). The framework imitates this procedure defining two types of episodic elements:
- The first one is related to the opportunistic sensing of the user’s whereabouts
- The second type is based on the virtual sensors of the phone
The research inspects and evaluates two different approaches for the generation of a SEM: one based on Finite State Machine generation; the latter leverages on frequent itemset mining and frequent sequence mining.
What will be the future work? “We want to lift the assumptions that the length of a SEM should be fixed to a day, enabling us to cover even longer SEMs. A second goal as future work regards the collection of a dataset that better represents the type of data handled by RoutineSense. […] Although the data collected by the RoutineSense are properly structured, additional research on semantic reasoning can be exploited to improve the capabilities of the framework.”
The paper will be published on EUDL European Union Digital Library.
