2. Activity Ads
People are interested in things they do
Use physical context to infer activity and determine
– Topics of interest
– Times when person is receptive to information
3. Activity Advertising
motivating vision
“An Inconven-
Work Transit Store Transit Dinner Transit Email Bed
ient Truth”
Today:
New Phone
Graham Crackers PDF Products
Plan
Activity Japanese
Targeted: “Bee Movie” Toyota Prius
Restaurant
PARC Confidential 3
4. Activity Stream Example Applications
… Activity … Activity … Activity … Activity … Activity … Activity … Activity … Activity … Activity … Activity … Activity … Activity … Activity …
Grocery Movie: “An
Work Transit Transit Dinner Inconvenient Truth” Transit Email Bed
Store
Target Save Minimize
Information Energy Waiting
• Determine the • Predict departures, • Predict transit
user’s needs and destinations, and route and time
interests arrivals • Notify to ensure
• Help advertisers • Optimize route to “just-in-time”
find receptive save fuel arrival at train or
consumers • Turn off power to meet a
when not in use colleague
PARC Confidential 4
5. Activity by Time of Day
how many people do what, and when
100% Miscellaneous
Traveling Traveling
90% Telephone Calls
performing each activity
80%
Percent of population
Volunteer Activiti
Socializing, Relaxing, Religious and Spir
70% and Leisure Sports, Exercise, a
60% Socializing, Relaxi
Education Eating Eating and Drinkin
50% and Government Serv
Sleeping / Household Servic
40% Work & Drinking Professional & Pe
Personal Care
30% Work-Related Consumer Purcha
Education
20% Work & Work-Rel
10% Household Activities Caring For & Help
Caring For & Help
0% Household Activities Household Activit
Personal Care
0 2 4 6 8 10 12 14 16 18 20 22
Hour of Day
This matches our intuition.
6. Activity Inference
a layered architecture
Name Data Sources Data Type Format Example
Venue Type, PhoneUse, Activity
Activity “Restaurant-ing”
FriendsActivities Taxonomy
Venue Venue Distribution, Type of Specific
“Restaurant”
Type SpecialPlacesList Venue
Location Distribution, “FukiSushi”=0.25,
Venue List of Venues
VenueDB, Accel, “PizzaChicago”=0.25,
Dist. & Probabilities
Calendar, Sound “SushiTomo”=0.5
lat=37.402, lon=-122.147,
Location Raw Position, GPS Coords +
Σ=[0.03, 0.01, 0.01, 0.04],
Dist. Accelerometer Uncertainty
time=145100
Raw Timestamped lat=37.402305, lon=-122.14769,
GPS
Position GPS Coords time=145107
PARC Confidential 6
7. Defining Activity
Taxonomy from ATUS 2006 (American Time-Use Survey)
Examples of the 18 Examples of the 110 Examples of the 462
Tier 1 Activities Tier 2 Activities Tier 3 Activities
Personal Care Sleeping Sleeping
Household Activities Grooming Sleeplessness
Caring For & Helping Household Members Health-related Self Care Sleeping, n.e.c.
Caring For & Helping NonHH Members Personal Activities
Work & Work-Related Activities Personal Care Emergencies Interior cleaning
Education Personal Care, n.e.c Laundry
Consumer Purchases Sewing, repairing, & maintaining textiles
Professional & Personal Care Services Housework Storing interior hh items, inc. food
… … Housework, n.e.c.
PARC Confidential
8. Time-Use Study Data
RESPID TIME ACTIVITY LOCATION
Physical care for Respondent’s home
20060101060033 07:00 - 07:20 household children or yard
Playing with Respondent’s home
20060101060033 07:20 - 09:20 children, not sports or yard
Physical care for Respondent’s home
20060101060033 09:20 - 10:20 household children or yard
Travel related to Car, truck, or
20060101060033 10:20 - 10:30 grocery shopping motorcycle (driver)
20060101060033 10:30 - 11:30 Grocery shopping Grocery store
ATUS 2006:
263,286 activity episodes 462 activities (Tier 3)
12,943 households 27 different location types
PARC Confidential
9. Activity Prediction Accuracy
for different sets of predictor variables
Percent Accuracy,
Percent Accuracy
Duration-Weighted Classifier
0%
0% 20% 40%
40% 60%
60% 80%
80%
None
Previous Tier 1 activity
Previous activity Tier 3
Tier 3
Previous activity & Day of week
Previous day Tier 2
Tier 2
Previous activity & Age Group
Previous activity & age group
Tier 1
Tier 1
Hour of day
Location and Time
Hour of day & Day of week
day of Day correctly
Hour of day & Age Group
Hour of day & age group
Hour of day & Day of week & Age Group
Hour of day & day of week & age group predicts activity
~60% of the time.
Previous activity & Hour of day
Previous hour
Location
activity & location
Previous activity & Location
& hour of
Location & Hour of day
Previousactivity & Location & Hour of day
Previous activity & location & hour of day
PARC Confidential
10. Activity Prediction Accuracy
at different locations
Percent Accuracy, Duration-Weighted Classifier,
Percent Accuracy
By Location
0%
0% 20%
20% 40% 60% 80%
80% 100%
Grocery store
Grocery store
Transportation
Transportation
Respondent's workplace
Respondent’s workplace
Gym, health club
Gym, health club At some locations,
Other store //mall
Other store mall
Bank
Bank
activity is predicted
Unspecified place
Unspecified place much better than 60%.
restaurant //bar
Restaurant bar
School
School
Someone else's home
Someone else’s home
At others,
Respondent's home
Respondent’s home it’s much worse.
Tier 3
1
Place of worship
Place of worship
Tier 2
Post office
Post office
Library
Library Tier 1
3
Outdoors away from home
Outdoors away from home Source: ATUS 2006
PARC Confidential
11. Predicting
Activities
from
Italian Chinese
Learned User
Patterns
Venue 50%
12:00 Likelihood: 50% 1:00
Weekly Behavior Patterns
Context History Monday Tuesda
Time Location Visit …
… …
11:57- 12:45 37°26’39”
12:00 $ $
-122°9’38”
to 1:00 $$ $$
1:22 - 1:31 37°23’11” Chinese Chinese
-122°9’02” Italian Italian
… … … …
1:00 to … …
12. Research Opportunities
in the advertising ecosystem
Ad Creator
user’s ad, bid, placement spec
predict activity? interest
ad receptivity? stream ad specification?
unfamiliarity? Interest Ad Network (e.g. optimal placement?
indeterminacy? Modeler Google) incentive balancing?
privacy modeling?
activity stream ad space details ad
GPS venue visit?
venue visit activity?
Activity Ad Space
reduce sampling needs? Inferencer Publisher
other sensors?
sensor data ad
When and where is best placement:
How to detect Finer-grained activities: Mobile display, ambient
Hobbies, exercise, sports, display, content sidebars, …?
vacation prefs,