I was fortunate enough to catch Duncan Watts' speech at ICWSM this year. I've taken a few notes on his presentation. I can't find the slides on the web, but below are the post-presentation notes that I took.

Initial discussion is about the telescope and its influence on science, and how social networks may be an equivalent tool for social sciences. Recurring theme is to leverage various social platforms to conduct social experiments that were infeasible (at best) prior to social networks. Coins the term "Social Science 2.0".

STUDY #1
Studied email network of college.

• Structural proximity overwhelmingly determines new ties.
• Proximate pairs tend to be similar.
• Individuals can meet via "cyclic closures" (intermediaries), and shared "foci" (focal closures) [not sure what focal closures are].
• "If we have 6 mutual friends and took 3 classes together, what is the probability that we know each other"

Findings seem to reflect what I've observed in my own studies. Also recurring theme with paper presented yesterday by Indika Kahanda (Using Transactional Information to Predict Link Strength in Online Social Networks).

STUDY #2
Experiments on Social Influence.

• Why are "hits" so much more successful than average?
• Very difficult to predict. Cites books and music.
• Implies unpredictability is derived from "social influence". Observation of others, and benefit of coordinated choices.
• What happens when everyone is influencing each other?
• Difficult to test both individual and collective at the same time, but required for valid conclusions.
• Used "Music Lab". Web-based "cultural market". Students shown a grid of 48 songs of unknown bands. Participants can rate and download the song. Participants are bucketed into two groups: independent, and social influence. Those in "social influence" bucket can see previous download counts for all songs. Social influence bucket significantly influenced by other listeners (as expected).
• Properties at collective level: Inequality of success (measured with Gini coefficient), and Unpredictability of success (average difference in market share of songs across R realizations of the "world" .. In this case, number of iterations of the experiment).
• Social influence worlds show more unpredictability and inequality than independent worlds. Best songs never do terrible, and worst songs never do very well, so world is not COMPLETELY unpredictable.

Stands in contrast to the Facebook Gesundheit! Paper (Gesundheit! Modeling Contagion Through Facebook News Feed) that was presented yesterday, which showed that page fanning in Facebook involved no social influencers.

STUDY #3
Network survey on Facebook.

• Hypothesis is that Americans are dividing themselves into like-minded groups.
• Other evidence shows the opposite. Reality and perception are different. People think their friends are similar to them, but they are not always so.
• Experiment: A's opinion about Q (a political question), B's opinion about Q, and what A thinks B's opinion is about Q.
• Very expensive to run this experiment (until the rise of social networks).
• Wrote "Friend Sense" Facebook App. Viral design proved to be quite effective.
• GSS-type study to get the data size from FB would have cost 200K-300K.
• Friends more similar than strangers, but not as similar as they think.
• Appears that respondents are simply unaware of what their friends think, rather than friends avoiding conflict.

STUDY #4
How do financial incentives affect performance?

• Performance based pay should elicit better work than a fixed wage.
• Focus instead on question: can an employer elicit better performance from a given pool of workers by paying them more?
• Leverage crowd sourcing to test pay scales.
• Study using Mechanical Turk.
• Subjects paid 10c per HIT (human intelligence task). Given a bonus if they correctly complete tasks.
• Assigned to 1 of 3 difficulty levels and 1 of 3 pay levels.
• Task is to order images taken over a 2 second interval.
• Mechanical Turk has an amazing diversity of participants (age/income/location). Well balanced gender breakdown as well.
• Finding: Subjects do more work for more pay. Increasing pay does not improve accuracy. Also, controlling for pay, people do less work for harder tasks.
• Participants perceived value of work shows "anchoring effect". Shows that people always think that they are underpaid. People paid 1c thought 5c, people paid 5c thought 8c, people paid 10c thought 13c.
• Tried same thing with word puzzles. Got similar results.
• Tentative suggestion is that payment levels should be dictated by recruiting and retaining efforts, not by work quality.

LINKS
ICWSM 2009
Gesundheit! Modeling Contagion Through Facebook News Feed blog comments powered by Disqus