1. Computing Trust in Social Networks
Jennifer Golbeck
College of Information Studies

2. Web-Based Social Networks (WBSNs)
Websites and interfaces that let people maintain browsable lists of friends
Last count
245 social networking websites
Over 850,000,000 accounts
Full list at

3. Using WBSNs
Lots of users, spending lots of time creating public information about their preferences
We should be able to use that to build better applications
When I want a recommendation, who do I ask?
The people I trust

4. Applications of Trust
With direct knowledge or a recommendation about how much to trust people, this value can be used as a filter in many applications
Since social networks are so prominent on the web, it is a public, accessible data source for determining the quality of annotations and information

5. Research Areas Inferring Trust Relationships
Using Trust in Applications

6. Inferring Trust
The Goal: Select two individuals - the source (node A) and sink (node C) - and recommend to the source how much to trust the sink.
tAC
PERSONALIZATION A B C
tAB
tBC

7. Methods TidalTrust SUNNY Profile Based
Personalized trust inference algorithm
SUNNY
Bayes Network algorithm that computes trust inferences and a confidence interval on the inferred value.
Profile Based
Trust from similarity

8. Source
Sink

9. Trust Algorithm
If the source does not know the sink, the source asks all of its friends how much to trust the sink, and computes a trust value by a weighted average
Neighbors repeat the process if they do not have a direct rating for the sink

10. Accuracy Comparison to other algorithms
Beth-Borcherding-Klein (BBK) 1994

11. Trust from Similarity
We know trust correlates with overall similarity (Ziegler and Golbeck, 2006)
Does trust capture more than just overall agreement?
Two Part Analysis
Controlled study to find profile similarity measures that relate to trust
Verification through application in a live system

12. Experimental Outline
Phase 1: Rate Movies - Subjects rate movies on the list
Ratings grouped as extreme (1,2,9,10) or far from average (≥4 different)
Create profiles of hypothetical users
Profile is a list of movies and the hypothetical user’s ratings of them
Subjects rate how much they would trust the person represented by the profile
Vary the profile’s ratings in a controlled way

13. Generating Profiles
Each profile contained exactly 10 movies, 4 from an experimental category and 6 from its complement
E.g. 4 movies with extreme ratings and 6 with non-extreme ratings
Control for average difference, standard deviation, etc. so we could see how differences on specific categories of films affected trust

14. Example Profile Movies m1 through m10 User ratings r1…r10 for m1…m10
r1…r4 are extreme (1,2,9, or 10)
r5…r10 are not extreme
Profile ratings pi = ri§i

15. Results
Reconfirmed that trust strongly correlates with overall similarity ().
Agreement on extremes ()
Largest single difference (r)
Subject’s propensity to trust ()

16. Extreme Ratings
When high are used on movies with extreme ratings, the trust ratings are significantly lower than when low are applied to those films
Statistically significant for all i

17. Maximum Difference (r)
Holding overall agreement and standard deviation constant, trust decreased as the single largest difference between the profile and the subject (r) increased.

18. Propensity to Trust ()
Aside from similarity, some people are more likely to trust than others
The average trust value a person has assigned () is a good predictor of how likely she is to trust a new person

19. Validation
Gather all pairs of FilmTrust users who have a known trust relationship and share movies in common
322 total user pairs
Develop a formula using the experimental parameters to estimate trust
Compute accuracy by comparing computed trust value with known value

20. In FilmTrust
Use weights (w1,w2, w3, w4, w) = (7,2,1,8,2)

21. Effect of change
If a node changes it’s trust value for another, that will propagate through the inferred values
How far? What is the magnitude? Does the impact increase or decrease with distance?
How does this relate to the algorithm?
Joint work with Ugur Kuter

23. Algorithms Considered
Eigentrust
Global algorithm
Like PageRank, but with weighted edges
Advogato
Finds paths through the network
Global group trust metric that uses a set of authoritative nodes to decide how trustworthy a person is
TidalTrust
TidalTrust++
No minimum distance - search the entire network

24. Initial ideas?
The further you get from the sink, the smaller the impact.
Changes by more central, highly connected nodes will create a bigger impact.

25. Network

26. Methodology Pick a pair of nodes in the network Repeat for every pair
Set trust to 0
Infer trust values between all pairs
Set trust to 1
Infer trust value between all pairs
Compare inferred values from trust=0 to trust=1
Repeat for every pair
Repeat for each algorithm

27. Fraction of Nodes at a Given Distance Whose Inferred Trust Value for the Sink Changed

28. Source
Sink

29. Average Magnitude of Change at a Given Distance

30. Conclusions and Future Directions

31. Conclusions Trust is an important relationship in social networks.
Social relationships are different than other common data used in CS research.
Trust can be computed in a variety of ways
The type of algorithm and behavior of users in the network impact the stability of trust inferences

32. Future Work - Computing with Trust
Major categories of trust inference: global vs. local, same scale vs. new scale
All have algorithms
Additional features (like confidence)
Hybrid approaches
Use trust assigned by users and similarity
Use multiple relationships for better certainty in certain domains (e.g. authority)

33. Future Work - Applications
What sort of applications can trust be used to support?
Recommender systems, filtering, tagging, information ranking
Disaster response
Highlight relevant items among vast collections of data

34. Jennifer Golbeck