Monday, 22 May 2006 1
Context Awareness – Edoardo Beutler & Sinja Helfenstein
Research Projects in the Area of Context Awareness
University of Zurich
Seminar Context Aware Computing
SS 2006
Sinja Helfenstein
Edoardo Beutler
Monday, 22 May 2006 2
Context Awareness – Edoardo Beutler & Sinja Helfenstein
Agenda
Context Aware Recommender Systems Sinja
COMPASS (COntext-aware Mobile Personal ASSistant)
UbiMate
Conclusion and Outlook
Recognizing the Places We Go Edoardo
BeaconPrint
Conclusion
Monday, 22 May 2006 3
Context Awareness – Edoardo Beutler & Sinja Helfenstein
Recommender Systems and Context Aware Systems
Recommender Systems
• Categorization & Recommendations based on interests / user profile
TripAdvisor.comFindAndDine.ch
Context Aware Systems
• Categorization & Recommendations based on user's current situation (context)
GPS Navigation Systems Location Based Services
Provision of information relevant to the user – in consideration of its context and interests
Static Environment!
Uniform Interests!
Goal: Provision of information relevant to the userInformation Selection Criteria:
• Hard: Filtering of useless items
• Soft: Rating of potentially useful items
Monday, 22 May 2006 4
Context Awareness – Edoardo Beutler & Sinja Helfenstein
COMPASSContext-aware Mobile Personal Assistant
Ratings
Buddies Interactivity:- Reservations- Calling buddies
Display as list or in map
Location-specific information for tourists:
Buildings, Restaurants, Hotels, Buddies, Taxi stands, Landmarks, etc.
Monday, 22 May 2006 5
Context Awareness – Edoardo Beutler & Sinja Helfenstein
COMPASS: Recommendation Criteria & Strategy
Recommendation Criteria:
• Hard: Current LocationSubscriptionsApplication Specifics
• Soft: Users Interests
Multiple Prediction Strategies possible, manual selection per item-group.
One Soft Criteria only One average rating per item Useful for Real World-information in a dynamic environment?
Example: Restaurant GuideBar Rimini:
clear sky, 35° = rainy, 10°?bqm:
with friends = with grandma?
[4] Human Factors Physical EnvironmentUser Social Env. Tasks Location Infrastruct. Phys. Cond.
Hard Subscript. Coordin.
Soft Interests Company
Open. hrs
Weather
Monday, 22 May 2006 6
Context Awareness – Edoardo Beutler & Sinja Helfenstein
Collaborative Filtering (CF)
Classical Collaborative Filtering
1) Compare users by their individual ratings and define neighbours (Pearson correlation coefficient)
2) Prediction for User u, based on neighbours' ratings(weighted by correlation)
Best-known example:
Monday, 22 May 2006 7
Context Awareness – Edoardo Beutler & Sinja Helfenstein
Introducing Context-Awareness to CF
• Association of various context-values to each rating
• Ratings' relevance for current prediction depending on:user AND context similarity!
• Implicit classification by context information in user feedback.
Advantages over classical CF-Systems:
Multiple ratings per item if used in different context.
Implicit feedback possible by inferring rating from user behaviour (e.g. duration of stay, frequency of visits).
Monday, 22 May 2006 8
Context Awareness – Edoardo Beutler & Sinja Helfenstein
[3] Mobile city guide based on context-aware collaborative filtering
• Method: Look at what like-minded user have done in the past under similar context to predict what the current user may like to do
• Context used: Information Source Hard/Soft
– User Information Manually defined user profile S– Social Environment Manually Entered by User S– Tasks (Activity) Manually Selected by User* H– Location GPS-Module H– Infrastructure -– Physical Conditions
• Weather Online Content Provider S
• Time Mobile Device S• *Currently available activities:
– Food (Restaurants, Bars, Take Away, etc.)– Entertainment (Sport, Culture, Spa, etc.)– Shopping (Groceries, Fashion, Art, etc.)
Monday, 22 May 2006 9
Context Awareness – Edoardo Beutler & Sinja Helfenstein
UbiMate : Demonstration
Testversion online:
http://ubimate.hopto.org
For Site Access:Username: friendsPassword: ubiubi
Personal registrationneeded for participation
Monday, 22 May 2006 10
Context Awareness – Edoardo Beutler & Sinja Helfenstein
Conclusion
• High potential for context awareness in mobile recommender systems
– Implicit feedback increases amount of ratings and data quality
– CF for handling the information flood resulting from multiple context dimensions (the more dimensions the better the prediction)
• Outlook– Improve context recognition and inference
– Improve usability
– Improve interactivity
Monday, 22 May 2006 11
Context Awareness – Edoardo Beutler & Sinja Helfenstein
BeaconPrint for Recognition of Places We Go
End goal: Define important places with names, not just coordinates.
Technique: WiFi and GMS, but not GPS (skyscraper-canyons)
BeaconPrint ...does: Provides a possibility to „extract“ relevant places from raw data.
“The mechanism for learning the physical destinations in someone'slife and detect whenever their devices return to those places“
does not: Assign automatically names or semantics to a place. (geocoding)
Monday, 22 May 2006 12
Context Awareness – Edoardo Beutler & Sinja Helfenstein
Algorithmic Tasks
Learning algorithm1. Segment assign a waypoint whenever the device is in a stable place.
2. Merge waypoints from repeat visits.
Recognition algorithm1. Recognize a device returning to a known place.
2. Recognize a device not in a place (mobile state).
→ BeaconPrint is a learning and recognition algorithm.
Monday, 22 May 2006 13
Context Awareness – Edoardo Beutler & Sinja Helfenstein
Related Work
Ashbrook and Starner's GPS Dropout plus Hierarchical Clustering AlgorithmMarking positions where for at least t minutes no GPS signal is received or the speed is below 1 mile per hour.
The comMotion Recurring GPS Dropout AlgorithmA position where the GPS signal is lost at least three times within a given radius is marked as important place.
Kang et al.'s Sensor-Agnostic Temporal Point Clustering AlgorithmAvoids the high dependence of a proper GPS signal by using temporalpoint clustering.
Monday, 22 May 2006 14
Context Awareness – Edoardo Beutler & Sinja Helfenstein
BeaconPrint Algorithm
• Gathering continually statistics about the radio environment.
• Parameters:Time window w - stable scans for at least w indicate a signifificant
place.Certainty parameter c in [0 ... c
max] and d = w/c
max - no new beacon
for d time indicates a stale scan.
• Not signal strength is the fingerprint metric, but constructs its fingerprint using a response-rate histogram (1-beacon loss rate responsrate).
Monday, 22 May 2006 15
Context Awareness – Edoardo Beutler & Sinja Helfenstein
Conclusion
• Runs on common hardware (WiFi, GSM)
• Recognizes and learns places to over 90% accurate.
• People have 72.3 places they go, only 1-2 frequent and 7-8 once a week.
• Former algorithms recognized only 5-35% of infrequent places (visited once for <10 min), BeaconPrint over 63%. In the second visit, the accuracy is increased to 80%.
Monday, 22 May 2006 18
Context Awareness – Edoardo Beutler & Sinja Helfenstein
References
[1] A. K. Dey, G. D. Abowd: Towards a Better Understanding of Context and Context-Awareness
[2] M. Van Stetten, S. Pokraev, J. Koolwaaji: Context-Aware Recommendations in the Mobile Tourist Application COMPASS
[3] A. Chen: Context-Aware Collaborative Filtering System: Predicting the User's Preferences in Ubiquitous Computing
[4] A. Schmidt, M. Beigl, H-W. Gellersen: There is more to context than location
[5] J. Hightower, S. Consolvo, A. LaMarca, I. Smith, J. Hughes: Learning and Recognizing the Places We Go