1
Thomas Sandholm, Hang Ung, Christina Aperjis, Bernardo HubermanHewlett-Packard, HP Labs, Social Computing LabRecSys, Barcelona September 27, 2010
GLOBAL BUDGETS FOR LOCAL RECOMMENDATIONS
2
WHY VOTE?
How do we get more people to contribute their opinions?
3
4
500USERS
30EXPERIMENTS
6LOCATIONS
Palo AltoChicago
Paris
Mumbai
Bangalore
Athens
5
MANY PEOPLE CONSUME CONTENT – FEW LEAVE OPINIONS
May 2008300M+
April 20108M
April 201027M
Sample takenUsers
6
RATING BUDGETS AND RANKING
Bob:$10
Alice: $11
John
http://dominos.com
Pizza
Restaurants
$4
http://burgerking.com
Burgers
Restaurants
$6
http://roundtable.com
Pizza
Restaurants
$3
http://applebees.com
Diners
Restaurants
$8
Top Channels
$8 Diners$7 Pizza$6 Burgers
Friend Channels
$6 Burgers$4 Pizza
$8 Applebee’s$6 Burger King$4 Dominos$3 RoundTable
7
RATING REWARDS
Bob
Alice
John
http://dominos.com
$4
Top Rewards
(reward factor 10)$40 Bob Dominos @ A$30 Alice RoundTable @ B
http://dominos.com
$5
$3 http://roundtable.com
http://roundtable.com
$3$3 http://roundtable.com
A
B
8
CLICK TO RATING RATIO
May 2008300M+
April 20108M
April 201027M
July 20104K
Sample takenUsers
9
RECOMMENDATION SUCCESS
Overa
ll
Cities
Hotel
s
Mov
ies
Mus
ic
Notes
Phot
os
Plac
esPo
lls
Resta
uran
ts
Shop
ping
Trav
el
Wikip
edia
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Clicks as Success
Ratings as Success
Success = proportion of query sessions ending with clicks/ratings
10
SYSTEM COVERAGE
11
MECHANICAL TURK EXPERIMENTS
Setup-5 Surveys-10 URLs-6 Locations-3 Continents-500 Users
Surveys-1-5 Star-Budget-Star Bonus-Budget Bonus-Gloe
12
MECHANISM RESULTS
1-5 Star Budget 1-5 Star Bonus
Budget Bonus
Gloe0
0.5
1
1.5
2
2.5
ParisPalo AltoMumbaiChicagoBangaloreAthens
Kendall Tau Rank Correlation (higher better)
1-5 Star Budget 1-5 Star Bonus
Budget Bonus
Gloe0
10
20
30
40
50
60
ParisPalo AltoMumbaiChicagoBangaloreAthens
RMSE (lower better)
13
BONUS EFFECT ON PARTICIPATION
2 3 4 5 60
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Positive SignalNegative Signal
Probability of signals
Number of surveys taken
14
LESSONS LEARNED
– Amazon Mechanical Turk•Workers not random geographic sample•Sensitive to task complexity•Respond well to small incentives
– Budget Mechanism•Higher quality recommendations with incentives•Social/Economic/Status value extract more opinions•Tuned based on usage, e.g. reward factor
15
FUTURE WORK
– Projects•AfricaMap: map annotation in remote parts for disaster relief UNOSAT/Uni. Geneva•Mobile print provider recommendations via HP ePrint
– Research•GSP Auction for commercial bidding•LMSR Market for recommendation arbitrage•Enhance reward mechanism to both encourage and identify high quality contributions
16
Papers at www.hpl.hp.com/research/scl
Live system at www.hpgloe.com
SOCIAL COMPUTING LAB
THANK YOU
1717
BACKUP
18
SYSTEM ARCHITECTURE
Clients
Java
JavaScriptjQuery
Android OS
BlackBerry OS
iUi/iPhone Safari
HTML5
Web Server
ApacheHTTP REST
Web Application Service
DjangoJSONP
GeoRSSHTML
Caching
Memcached
Integration
Services
Google MapsGeocoding
FacebookSocialAuth
OAuth
Database Shards
MySQLGeohashing
19
SCALABILITY
20
HP GLOE: STATUS
~7m* recommendations at http://hpgloe.com
~4k* users on Android, iPhone, BlackBerry, WebOS, Web…
*Sept 2010
21
LESSONS LEARNED CONTINUED
– Gloe system•Geohash location partitioning simple and efficient•HTTP(S) GET/JSON(P) has served us well on all platforms•MySQL & Sharded architecture flexible and fast