CSCI 8970 – Colloquium Series – Fall 2010 – Ninth Event
Exploring the Structure of On-Line Social Networks: The Role of Positive and Negative Interactions
Cray Distinguished Speaker Series
Monday, November 1, 2010
|
Presenter: |
Jon Kleinberg, Cornell University |
Is there a relationship between friends and our relationships with others? What can be extracted from social networks and the newly accessible data they have accumulated regarding people’s interaction? Jon Kleinberg, a well known scholar in computer science is trying to answer this question by looking at variables such as reputation, recommendation and ranking. Things were previously invisible are now leaving traces. Captivated by Stanley Milgram’s six degrees of separation study were letters were at randomly dispersed in Indiana to see how many of them reached Boston (1/3rd actually made it). Where he came up with an estimated distance or optimal exponent of 2.Now with data from online social networks such as Facebook and LiveJournal research resulted in 1/rank (v,w)^1.05 or 1/rank (v,w)^0.95 and a distance of d –> rank – d^2. The data, using two different data sets provided similar results.
After concluding this study, Kleinberg has since decided to revisit the great challenges of 21st century sociology. “Why is it that you are similar to your friends? Because they influence you, or because you seek out people who are already similar?” Having a large number of previous answers, Kleinberg looked at social network data to find an answer. Current data sets tend to be more supportive of positive rather than negative link research. Most links made in places such as Facebook have positive connotations, yet some websites provide insight into negative relations such as Epinions (Trust / distrust) and Slashdot (Friend/Foe).
Yet, more generally, users express positive and negative attitudes implicitly, through various kinds of actions, including voting for admin promotion on Wikipedia {Burke-Kraut 2008], and possibilities for researching multi-player on-line games [Szell et al. 2010]. In conducting their research, Kleinberg and his team focused on two competing theoretical frameworks: theory of balance and theory of status. According to Balance Theory [Heider 1946, Cartwright-Harary 1956] there are 16 different relationship triangles, yet the ones that make sense are the ones that mention friends an odd number of times such as: the friend of my enemy is my enemy, the friend of my friend is my friend. Things like the friend of my friend is my enemy – tend to resolve themselves, one of them pulls his way against the other.
According to Status Theory, traditional methods of adapting balance theory: disregard directions; apply undirected formulation [Wasserman-Faust 1994]. Three friends, do not simply all look at each other as equals. C could recommend B, B could recommend A, but A politely may not recommend C.
In their results there were both aspects of Balance and Status theory available in the data. Unfortunately, grouping all positive links under the term friend limits a more specific study. He gave an example of soccer players to more clearly explain status theory. Having three players responds. While X gives good ranking to A and B (which are above average players), and A and B give out a lot of evaluations, and both are considered good players, one is positive toward the other, while the other one is negative towards the other. These results proved consistent with 27 out of 32 possibilities having the directionality correct.Mutual link are rare (5%), but one dimensional ranking links (status based) are more common
Wikipedia – Admin promotion analysis
- You do have number of articles edited, awards (Barnstars), etc.. (CV of work)
- 100,000 votes for promotion
- Probability of a positive vote analysis – (a conjectured plot)
- Voters are particularly harsh on the individuals that have the same level of achievement that they do.
- Competition with these person.
- You can also see the commonalities between them more easily.
- They are also doing this cross culturally – German (a bit different), French (similar)
Learning Links Signs from Network Context
- Based on triad types – status and balance theory are just two general theories
- Via logistic regression, they can learn a theory of link signs.
- Accuracy 80-93% when placing a model from one area of Wikipedia into another.
Conclusion
Analysis of signed networks provides insights into how social computing application are being used
Status and balance
Rich connections between learn models and classical theories
Use of edge signs depends on embeddedness, and reciprocation
Approximate versions of the model
Different sign patterns produce different amounts of status and balance: an “energy landscape” [Marvel-Strogatz-Kleinberg 2009]
Many further directions and open questions
