Dawn Nafus's presentation at eComm 2008
Post on 15-May-2015
431 Views
Preview:
Transcript
Dawn Nafus
Tye Rattenbury
Ken Anderson
Add GPS and Stir? Some Context for Context Awareness
2
The shift to “context aware” systems has been a shift away from active user input toward leveraging passively collected datasets Systems now add value by giving new life to data
– Nike uses accelerometer data to tell customers about their run– Google takes search terms to ‘contextualize’ ad placement– Social networking sites present new contacts, ‘community’ visualizations
based on system data
Though computer scientists would have a stricter, more additive, definition -
Accelerometer + Humidity= running contextNike would have to add in humidity measurements in order to infer something about the run
Person + Location =Friend
Person + Location+ Same MP3s on iPhone=Potential New Friend
3
There’s excitement both in the marketplace AND in labs
Like Bill Murray’s Groundhog Day,
some tech scenarios just
won’t go away.
On the market•Google’s astronomical rise through ‘contextual’ ads•Other opportunities persist as the perpetual ‘next big thing’
•The ‘use location data to throw coupons at the passerby’ model won’t die•Many new friend finding startups, though not yet mass market
In the lab•Sensor technology rapidly developing•Inferencing and machine learning capabilities are on the increase•Compute resources growing (data aggregation and sharing, compute power)
4
… but there’s something fishy about the word ‘context’
Sensor context
Lighting
Noise level
Location
Movement
Other devices
Other people
Things that are machine readable
People context
Emergent action
What’s relevant to the conversation
Blink vs wink
House vs home
Commute vs journey
Cooking for pleasure vs cooking for food intake
Things that are person readable
Source: P. Dourish, “What We Talk About When We Talk About Context.” Ubicomp 2004
ROI(getting
people to value what you have)
Culture(how people understand the world
around them)
5
Which context you focus on leads you to different questions, which lead to different design choices and business models
How much can I reliably detect?
How many different data points can I throw together?
How can I make machines mash it all up?
What activity can I support?
How do my end users interpret my data?
Sensor Context
Things that are machine readable
People Context
Things that are person readable
6
Mobile Friend Finding vs the Whereabouts Clock
Sensor Context: Location + buddylist=friendfinding
People Context: relative to no one in particular
Sensor Context: Location +buddylist =ambient display of routine
People Context: Relative to spouses and children
Embedded in routines of ‘putting on the kettle’, practices of assurance-giving, as precise as conversations are, allows redefinition of home and away
“Whereabouts Clock”, Brown et al 2006
Location ≠ Context
7
Think very, very carefully about the following question.
8
The average office worker has 12 minutes to work before he/she gets interrupted.
You now have a device that tells you whether people of potential relevance to your work are in the office, based on location, other devices present, and historical interaction pattern.
Do you really want something to interrupt you to tell you that you are about to be
interrupted?
9
Context-aware tourism applications
Paris photo courtesy of Only_Point_Five
•Easily machine sensed•People context present, ritualized•low demand for machine-inferred Eiffel
Towers
•Too many things to ‘sense’ (building? Shop? Construction materials?)•People context present, but unritualized (unpredictable)•High value– the ‘gem’ you discover on your trip!
• Machine recognizable but (semi)long-tail. Who will connect the dots?•Value depends on people’s enthusiasm for a narrow art genre, not necessarily a particular place
Easy, low value Hard, high value
Middling
10
Spurious connections between Sensor Context and People Context are fantastic for art, but cloud ‘prediction’
“Home Health Horoscope” Gaver et al, 2007Sensor Context:Condensation on windows + pattern of doors + weight of the coffeepot + … = home health
People Context:Sensors mash up horoscopes and give them back to perplexed dwellers who infer their meaning (i.e., do the real ‘sensing’ work)
Inferencing that involves more than 10 variables will deliver spurious connections.
What will people make of these?
11
Current research: Mapping connections between People Context and Sensor Context
•Data from 8 weeks of device activity used as a cultural probe
•How the device ‘fits in’ to people’s daily patterns says a lot about both their lives and what they believe the device ‘does’
•Key result: machine use is the aspect of life most likely to be interrupted, doesn’t do the interrupting
12
With a weak ecosystem of players focused on people context, we all suffer sustainability problems
• Places for eyeballs to go become limited if no one interacts with the device
• Creepyness factor becomes the foreground– People need a good reason to not notice they are creating a database for
you
• Bias towards literal accuracy can be more annoying rather than less
• Non-traditional ads are by definition ‘out of place’—the bar is raised to get it right
13
If Location ≠ Context, then…
• “Context awareness” as a technical development obliges businesses to do MORE user-centered design, not less– Machine learning will only solve the problems you design it to solve– You don’t need an anthropologist to do this, just some good observations about daily life
• Use people context to assess the value of sensors AND constrain the noise– Understanding how the two map on to each other helps you understand what people
value– A ‘less is more’ approach to data is pretty useful before full AI happens
• When you do it right, you’ve done something pretty powerful– you are creating a system of meanings that people use to perceive their world– this is
culture
15
Orkut in Brazil: where People Context made Sensor Context mean something (accidentally)
In January 2005, Brazilians recruited masses of friends onto the social networking site Orkut in order to “beat the Americans”.
– Math mattered unexpectedly: People felt they “had been counted” in a people context where they felt invisible (US-dominated Internet)
– Nothing about social network density, size or shape could have inferred that a nascent social movement was brewing. Social history would have.
– “Takeover day”-the day when the number of Brazillians was greater than Americans- was a national event
this became ‘we’
This did not
An eComm 2008 presentation –
http://eCommMedia.com for more
top related