SIGIR 2007 Proceedings

Doctoral Consortium

Information-Behaviour Modeling with External Cues
Michael Huggett
University of British Columbia 2366 Main Mall Vancouver, Canada V6T 1Z4 Tel: 604 822-3061

mike@cs.ubc.ca ABSTRACT
Much of human activity defines an information context. We awaken, start work, and hold meetings at roughly the same time every day, and retrieve the same information items (day planners, itineraries, schedules, agendas, reports, menus, web pages, etc.) for many of these activities. Information retrieval systems in general lack sensitivity to such recurrent context, requiring users to remember and re-enter search cues for objects regardless of how regularly or consistently the objects are used, and to develop ad-hoc storage strategies. We propose that in addition to semantic cues, information objects should also be indexed by temporal and sensory cues, such as clock time and location, so that objects can be retrieved by external environmental context, in addition to any internal semantic content. Our cue-event-object (CEO) model uses a network representation to associate information objects with the times and conditions (location, weather, etc.) when they are typically used. Users can query the system to review their activities, revealing what they do at particular times, and which information objects tend to be most often used and when. The system can also pre-fetch items that have proven useful in past similar situations. The CEO model is incremental, real-time, and dynamic, maintaining an accurate summary even as a user's information behaviour changes over time. Such environmentally-aware systems have applications in personal information management, mobile devices, and smart homes. As a memory prosthesis, the model can support autonomous living for the cognitively impaired. We present a comprehensive research agenda based on some promising preliminary findings. objects such as descriptive keywords. Links can represent many different types of relation such as is-a or has-component, but complex networks with many link types often need to be built manually due to their semantic subtlety [2]. By contrast, the task of building semantic networks from large corpora is too timeconsuming to be done manually. The simplest automaticallyconstructed networks use just a single weighted link type that represents the degree of relatedness between objects. Relatedness can be calculated automatically using a similarity function to compare objects, and objects are retrieved through a traversal process such as spreading activation. [1, 3] The Cue-Event-Object (CEO) model is an online real-time system that automatically builds an incremental representation of the typicality of recurring events. Its cues represent stimuli external to the system, such as time, location, temperature, velocity, luminance, etc. Events represent a particular type of occurrence, such as "going to the opera", and are linked to cues that indicate the circumstances of the event. An object is a datum linked to the event, such as "concert program" or "menu". A single event type may be linked to multiple cues and objects. Links grow stronger as events re-occur under repeated circumstances, to create emergent patterns. This enables such queries as "what do I usually do at time T?", or "when do I usually do activity A?" The CEO model has been tested in a preliminary evaluation that uses a 6-month sequence of user behaviours to predict future actions. Using a very simple implementation with basic time and location data, the model learned patterns quickly, and gave 83% accuracy of prediction. Further evaluation will measure the reliability of patterns, the ability to capture their "drift" as user behaviour changes, the effect of increasing the number of cues, and the model's utility with mobile sensors. Longitudinal studies will also be used. Thus we will show how external cues can complement standard semantic indexing to retrieve information objects under repeated conditions, such as at scheduled meetings, or en route to particular destinations. The CEO model can be used to add context dependence to any information system that is organized in terms of discrete information objects.

Categories and Subject Descriptors
H.3.3 [Information Search and Retrieval]: retrieval models

General Terms
Algorithms, Experimentation, Human Factors.

Keywords
Contextual retrieval, information behaviour, sensor indexing, spreading activation networks, temporal indexing

References
[1] Belew, R.K. Adaptive Information Retrieval. In Proceedings of SIGIR, 1989: 11-20. [2] Cohen, P.R. and Kjeldsen, R. Information retrieval by constrained spreading activation in semantic networks. Info..Proc. and Management, 23 (4). 255-268, 1987. [3] Jones, W.P. The Memory Extender Personal Filing System. In Proceedings of SIGCHI, 1986: 298-305.

The CEO Model
Semantic networks are a popular way to organize information and display its relationships. Nodes represent documents or other information objects, and often also represent attributes of the

Copyright is held by the author/owner(s). SIGIR'07, July 23­27, 2007, Amsterdam, The Netherlands. ACM 978-1-59593-597-7/07/0007.

920