U NIVERSITY OF M INNESOTA Altruism, Selfishness, and Destructiveness on the Social Web GroupLens...
-
Upload
josiah-bann -
Category
Documents
-
view
215 -
download
1
Transcript of U NIVERSITY OF M INNESOTA Altruism, Selfishness, and Destructiveness on the Social Web GroupLens...
UNIVERSITY OF MINNESOTA
Altruism, Selfishness, and Destructiveness
on the Social Web
GroupLens ResearchUniversity of Minnesota
John Riedl
UNIVERSITY OF MINNESOTA
Bowling Alone (Amazon reviews)
UNIVERSITY OF MINNESOTA
Adaptive Hypermedia 20084
Tags scale:• Library of Congress: 20M books in 200
years.• www.librarything.com: 22M books in 3
years.Tag draw relevance from “the wisdom of crowds”
Adaptive Hypermedia 20085
Messages
Community-maintained Artifacts of Lasting ValueoRequires User Modeling and Adaptive
Hypermedia
Key Research Challenges:oAttract contributionsoMaintain qualityoAchieve agreement
Adaptive Hypermedia 20086
Alexa Germany
UNIVERSITY OF MINNESOTA
1. Google (German)3. Google (English)
Search
Adaptive Hypermedia 20088
Adaptive Hypermedia 20089
Google PageRank
Value of a page is the value of the pages that link to it
Recursive!
Algorithms and PsychologyThe Rich get Richer
Adaptive Hypermedia 200810
Web Structure
UNIVERSITY OF MINNESOTA
(Web Search)shared
Maurice Coyle and Barry SmythAH’08
Adaptive Hypermedia 200812
Research Questions
How can we mine free activity?What are the risks in these data?
UNIVERSITY OF MINNESOTA
2. YouTube
Video by Amateurs
Adaptive Hypermedia 200814
Chocolate Rainby Tay Zonday
Adam Bahner, a Ph.D. student in American Studies at the University of MinnesotaNumber 2 hottest viral video in historyoHottest viral video of Summer 2007oOver 26 million views
Adaptive Hypermedia 200815
Videos Life Fast, Die Young
Adaptive Hypermedia 200816
Adaptive Hypermedia 200817
Huberman Dynamics of Viral MarketingThe Dynamics of Viral Marketing,
ACM TWeb 2007, Leskovec et al., HP
Adaptive Hypermedia 200818
Maximizing the Spread of Influence through a Social Network, David Kempe, Jon Kleinberg,
Éva Tardos, KDD’03
Independent Cascade Modelo Information diffuses over timeo Each neighbor who converts has a
one-time chance to convert others
Linear Threshold Modelo Each node considers the preferences
of all neighborso If total weight passes threshold, a
node converts
Adaptive Hypermedia 200819
Video suggestion and discovery for YouTube: Taking random walks through the view graph
Shumeet Baluja, et al., Google, WWW 2008
Adaptive Hypermedia 200820
Research Questions
How do preferences propagate naturally?What predicts fads?How do recommenders influence propagation?
UNIVERSITY OF MINNESOTA
4. Ebay
Online AuctionsCustomers Selling to Customers
Adaptive Hypermedia 200822
Google Trends Front Page
Adaptive Hypermedia 200823
Adaptive Hypermedia 200824
4Chan vs. eBaumsWorld
4ChanoGoogle Trends HackoChocolate Rain
eBaumsWorldoMany other hackso “copyright” fight with 4chan
UNIVERSITY OF MINNESOTA
UNIVERSITY OF MINNESOTA
The Internet is Serious Business
“A phrase used to remind those who voluntarily leave the house that being mocked on the Internet is, in fact, the end of the world.”- Encyclopedia Dramatica
Adaptive Hypermedia 200827
Amazon Robertson
shilled
Adaptive Hypermedia 200828
The Information Cost of Manipulation-Resistance in Recommender Systems Resnick and Sami. ACM RecSys 08.
The Social Cost of Cheap PseudonymsFriedman and Resnick, Journal of Economics and Management Strategy, 2001
UNIVERSITY OF MINNESOTA
Increasing Contributions
Adaptive Hypermedia 200830
What Theory Tells Us…Collective Effort Model People will contribute more if:
They believe their effort is important to the group
They like the groupSmaller is Better Slovic, Fischhoff, & Lichtenstein, 1980 People feel greater concern when the
reference group they’re part of grows smaller.
Specificity Matters Small & Loewenstein, 2003 Specific identity of those helped is
important in drawing people’s support.
Adaptive Hypermedia 200831
CommunityLab Research
Social science to increase contributions Accessible to designers Algorithms, interfaces, toolkits
GroupLens @ Minnesota Recommender algorithms and
interfaces John Riedl, Joe Konstan, Loren Terveen
Bob Kraut and Sara Kiesler @ CMU Social psychology of computer use
Paul Resnick and Yan Chen @ Michigan
Adaptive Hypermedia 200832
VOICE 2 Screen shotNumerical values are represented
by smilies
Who the contribution helps
Value of each contribution
Adaptive Hypermedia 200833
Results
Want Smilies on the regular interface?
Self-report
1: Strongly Disagree2: Disagree3: Neutral4: Agree5: Strongly Agree
1 2 3 4 5
Self 3.87
All MovieLens 3.13
Similar Group 2.97
Dissimilar Group
2.94
Control 2.68
0% 5% 10% 15% 20%
Probability of rating a movie
Behavioral data
Self 7.2%
All MovieLens 10.2%
Similar Group 15.8%
Dissimilar
Group 5.9%
Control 7.4%
Adaptive Hypermedia 200834
Research Questions
How can contributors be motivated?How can social attacks be mitigated?oMail list “unsubscribe”
How does social psychology interact with defense algorithms?oCan the griefers be encouraged to
give up?
Can freedoms be preserved?
UNIVERSITY OF MINNESOTA
5. Yahoo!
Everything
Adaptive Hypermedia 200836
Flickr Popular Tags
Adaptive Hypermedia 200837
Tag Selection Algorithms
“The Quest for Quality Tags”S. Sen, F. Harper, A. LaPitz, J. RiedlGROUP 2007
Adaptive Hypermedia 200838
Catcher in the Rye
Huge number of tagsRQ: How can a tagging system show users tags
they want to see?
Adaptive Hypermedia 200839
Users don’t agree
Most controversial tags (Bayesian expected entropy):tag entropy # #
comedy 0.987 28 30
classic 0.986 25 24
stylized 0.983 20 21
nudity (full frontal) 0.980 18 20
romance 0.980 18 17
quirky 0.977 25 20
magic 0.974 18 15
animation 0.974 26 20
Steven Spielberg 0.973 12 12
sci-fi 0.972 14 17
Adaptive Hypermedia 200840
Tag PredictionRandom baseline: 21%
Implicit features:number of applications (39%)number of users (51%)number of searches for a tag (44%)number of users who searched for a tag (48%)length of tag (42%)
Moderation-based features:global average rating for a tag (59%)user-normalized global average rating for a tag (62%)tag reputation (57%)
Hybrid combinations: logistic regression, decision trees (67%)
Adaptive Hypermedia 200841
Research Questions
How can a system distinguish between “good” tags and “bad” tags?How should quality control work?Can folksonomy be encouraged? o Showing users more tags leads to more
vocabulary reuse oHow much convergence is valuable?
UNIVERSITY OF MINNESOTA
6. Wikipedia
Next slide, please!
Adaptive Hypermedia 200843
Wikipedia on Wikipedia
UNIVERSITY OF MINNESOTA
Wikiality on MySpace
1:20 – 2:15: edit wikipedia to make truth“What if the number of elephants in Africa were increasing?”
UNIVERSITY OF MINNESOTA
Creating, Destroying, and Restoring Value in
WikipediaGroup 2007
Reid PriedhorskyJilin ChenShyong (Tony) K. LamKatherine PancieraLoren TerveenJohn Riedl
Adaptive Hypermedia 200846
Adaptive Hypermedia 200847
Adaptive Hypermedia 200848
Adaptive Hypermedia 200849
Who contributes Wikipedia’s value?
User:Maveric149
3.8 million least frequent
editors0.5% of value 14% of valueWales
Swartz
Adaptive Hypermedia 200850
PWV contributions of elite editors
Adaptive Hypermedia 200851
Adaptive Hypermedia 200852
Adaptive Hypermedia 200853
Research Questions
How can vandalism be detected?How efficient is Wikipedia?How much conflict is valuable?
UNIVERSITY OF MINNESOTA
7. Studiverzeichnis
Social Network
Adaptive Hypermedia 200855
Adaptive Hypermedia 200856
Adaptive Hypermedia 200857
Adaptive Hypermedia 200858
The Predictive Power of Online Chatter
• Gruhl, Guha, Kumar, Novak, Tomkins
• Yahoo• ACM KDD 2005
• Volume of blog postings predict sales rank of books
• Queries can be automatically generated in many cases.
• Can sometimes predict spikes in sales rank.
Adaptive Hypermedia 200859
Anti-aliasing on the Web
Jasmine Novak, Prabhakar Raghavan, Andrew Tomkins.
WWW 2004
Adaptive Hypermedia 200860
ZipBirthdat
eSex
Story: Finding Medical Records (Sweeney 2002)
Medical Data EthnicityVisit DateDiagnosisProcedureMedicationTotal Charge
Voter List NameAddressDate
registeredParty
affiliationDate last
voted
ZipBirthdate
Sex
Former Governer of Massachussetts!
Adaptive Hypermedia 200861
Risk of Information Exposure (Frankowski et al., SIGIR ‘06)
Sparse Dataset 1: private
YOU
Sparse Dataset 2: public
YOU
+ +
= Your private data revealed!
Combining algs
Keep private information within domain!
Adaptive Hypermedia 200862
MovieLens Forums
- Started June 2005
- Users talk about movies
- Public: on the web, no login to read
- Can people identify these users in our anonymized dataset?
Adaptive Hypermedia 200863
Research Questions
Can users be identified from the personal recommendation data? YESCan the datasets be redacted to protect the users? UNKNOWNCan the users be warned in time? OPEN QUESTION
Adaptive Hypermedia 200864
Adaptive Hypermedia 200865
Adaptive Hypermedia 200866
Quantity
Quality
Tags
Social
Identity
Research Practice
Concept Understanding
Adaptive Hypermedia 200867
Messages
Community-maintained Artifacts of Lasting ValueoRequires User Modeling and Adaptive
Hypermedia
Key Research Challenges:oAttract contributionsoMaintain qualityoAchieve agreement
Adaptive Hypermedia 200868
Acknowledgements GroupLens
o John Riedl, Joe Konstan, Loren Terveen o Dan Cosley, Shilad Sen, Tony Lam, Rich Davies, Dan Frankowski,
Max Harper, Sara Drenner, Al Mamunur Rashid, Sean McNee, Reid Priedhorsky, Aaron Halfaker
CommunityLabo Sara Kiesler, Bob Kraut, Paul Resnick, Yan Chen
NSFo DGE 95-54517, IIS 96-13960, IIS 97-34442, IIS 99-78717, IIS 01-
02229, IIS 03-24851, IIS 05-34420, IIS 03-25837