Computing and Applying Trust In WebBased Social Networks

1
Computing and Applying TrustIn Web-Based Social
Networks

• Jennifer Golbeck
• August 18, 2005

2
Web-based Social Networks(WBSNs)

• 135,000,000 user accounts over 127 different
  networks
• 18 sites with over 1,000,000 members
• 6 sites with over 10,000,000 (Tickle, Friendster,
  Adult Friend Finder, Black Planet, Hi5, MySpace)
• Social/Entertainment, Dating, Religious, Photos,
  Pets, Blogging, Business
• 7 sites have FOAF output

3
Motivations

• Public interest
• Emergence of social trust in computer science
• Quantifying and computing in social networks is
  hard - concepts are fuzzy

4
Trust in WBSNs
5
About Trust

• To make computations about trust, we must
  understand the properties of the relationship
• Trust is fuzzy concept, and is being expressed in
  a social way. The definition and properties are
  not mathematical formalisms, but social ones.

6
Definition

• Two components
• Belief
• Commitment
• Definition Alice trusts Bob if she commits to an
  action based on the belief that Bobs actions
  will lead to a good outcome for her

7
Properties

• Note subject specificity
• Transitivity (non-mathematical)
• Composability
• Asymmetry
• Personalization

8
Inferring Trust in WBSNs
9
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
A
B
C
tAB
tBC
10
Inferring Trust in WBSNs Continuous Values
11
Continuous Trust Ratings

• Do trusted people tend to agree with us more?
• How does distance affect agreement/accuracy?

12
Trust Project Network
13
FilmTrust Network
14
Trust and Length
15
TidalTrust An Algorithm for Inferring Trust

• Breadth First Search based search from source to
  sink
• To improve accuracy
• Search minimum possible depth
• Accept ratings from only the highest rated
  neighbors
• Weighted average
• Modify based on analysis of specific networks

16
Complexity and Accuracy

• Complexity O(VE) - basically O(E) for our
  networks, where EV
• Comparison to other algorithms
• Beth-Borcherding-Klein (BBK) 1994

17
Applications and Extensions

• Trust and Social Networks in Email

18
Building Social Networks from Email

• Extract threads from corpus
• Identify threads related to the project (NLP
  technique)
• Use an email between people as a connection on
  the date when it was sent

19
TrustMail
20
TrustMail

• Filter email messages according to the trust
  value of the sender
• Expect high coverage if users rate the people to
  whom they send messages
• Analysis of Enron Email Corpus suggests about 92
  of senders could be rated (and an even higher
  percentage of messages).
• 37 of messages were from people the recipient
  had emailed
• 55 were from people who could be found through
  paths in the social network

21
Relationship Dynamics in Social Networks

• Over the course of a project, how does the
  social network change?
• Look at all people involved in a project over its
  life, and the connections between those people
• Theories
• Core group involved in the entire lifecycle
• Different clusters enter and leave the network at
  different times

22
Applications

• Detect groups that are involved with projects at
  different stages
• E.g. If a group is active in the early stages of
  a project and seen active again, it may signal
  another project has started
• Identify the core group by noticing its presence
  while other clusters move in and out
• Overall pinpoint the roles of groups of people
  within the network.

23
Temporal Social Networks from Email

• Modify the strength of a connection based on how
  the person was addressed (direct To, CC, BCC,
  etc)
• Decay in strength of relationship over time

24
Visualization
25
Modes of analysis

• Identification of clusters
• Difficulty increased with temporal aspects
• Determining network centrality

26
Measuring Centrality

• Centrality can be computed with several measures
• Over time, how does the centrality of nodes
  change?

27
(No Transcript)
28
Current Work

• Manually identifying projects within the Enron
  email collection
• Adapting existing social network algorithms for
  centrality and clustering to temporal networks
• Developing algorithms for decay of social
  connections
• Testing hypotheses
• Other projects can be identified by active
  clusters
• Central clusters can be identified successfully

29
Conclusions and Future Work
30
Conclusions

• Trust can be inferred in social networks with a
  good degree of accuracy
• These inferences can be incorporated into
  applications to offer usability benefits
• Similar techniques on other relationships may
  hold promise for gaining insight into social
  networks.

31
More Information

• http//trust.mindswap.org
• golbeck_at_cs.umd.edu

32
(No Transcript)
33
FilmTrust

• http//trust.mindswap.org/FilmTrust
• Combines online social network (w/trust) with
  movie ratings and reviews
• Use trust inferences
• To customize ratings
• To sort reviews

34
(No Transcript)
35
(No Transcript)
36
(No Transcript)
37
Average Error (?) in Predictive Ratings
(low is better)
38
FilmTrust

• http//trust.mindswap.org/FilmTrust

39
Inferring Trust in WBSNs Binary Values
40
Binary Valued Networks

• Users state their relationship with neighbors as
  having trust or no trust
• Values can be represented as 1 for trust and 0
  for no trust

41
Inferring Trust in Binary Networks

• Breadth First Search
• Source asks trusted neighbors for their opinions
  of sink, takes the average of their responses.
  Round to get a 0 or 1 value (recursive)

42
Determining Accuracy

• Know the sources opinion of every node
• Let g be the percentage of good nodes (who
  honestly try to provide good information).
• Let pa be the accuracy of good nodes (the
  probability that they agree with the source)
• Let a gpa

43
Calculating Accuracy
Probability that a majority of neighbors will
return the correct value
Probability of a correct result as n increases,
for a0.5
Detail
44
Experimental Data

• Generated networks to see the effect of changing
  the parameters
• Used Watts-Strogatz ß-graph model to create
  networks with small world structure
• Networks with 400, 1000 nodes
• Use g and pa to assign trust ratings to edges
• Vary (g,pa) by 0.05. For each pair, inferred
  values on 1,000 networks
• Result if a 0.5, accuracy remains high

45
(No Transcript)
46
Backup/Additional Slides
47
Hypothesis

• Trust in web-based social networks can be
  computed and integrated into applications to
  enhance the users experience.

48
Contributions

• Developed a formal definition of trust as a
  computational concept within WBSNs
• Developed algorithms for inferring trust
  relationships and demonstrated they are highly
  accurate
• Showed that trust-based predictive
  recommendations outperform other methods when the
  users opinion is divergent from the average

49
Trust in Application

• Whats New Showed that trust-based predictive
  recommendations outperform other methods when the
  users opinion is divergent from the average

50
Relationships in WBSNs

• 54 sites allow users to say things about their
  relationships
• 6 sites let users express trust relationships

51
What is a Web Based Social Network (WBSN)?

• Accessible over the web with a browser
• Users explicitly state their relationships to
  others qua stating a relationship
• Relationships must be visible and browsable by
  other users in the system
• There must be explicit, built-in support for
  creating these relationship statements

52
(No Transcript)
53
Future Work Validation

• To verify these results, the algorithms must be
  applied in much larger networks
• E.g. Integration of Gmail and Orkut

54
Future Work Extensions

• Further analysis of the relationship of network
  structure, algorithmic accuracy, and dynamics
  within the network
• Building a better recommender system - how trust
  can be integrated into traditional ACF methods
• Application of trust to other problems filtering
  statements on the semantic web

55
Broader Future Work
56
Other Work (not mentioned here)

• Verification of the meaning of trust in
  networks
• Network visualization
• Thorough analysis of the effect of trust rating
  and path length on accuracy in continuous
  networks
• Analysis of Enron Email Corpus for TrustMail

57
Calculating Accuracy
Probability that a majority of neighbors will
return the correct value
Probability of a correct result as n increases,
for a0.5
58
(No Transcript)
59
Trust and Accuracy

• Are trusted people more accurate?

60
(No Transcript)
61
Path length and Accuracy
62
(No Transcript)
63
Trust in WBSNs

• Orkut (1-3)
• Trust Project and FilmTrust (1-10)
• Rep Check
• Business Trust and Personal Trust (1-5)
• Overstock Auctions
• Business Rating (-2 - 2)
• Personal Rating (0-5)

64
Networks for Testing

• To test the accuracy of these algorithms, we
  generated small world networks that are accurate
  models of social networks
• Two Variables
• Good Nodes vs. Bad Nodes (g)
• Accuracy of Good Nodes (pa)

65
Expected Values
66
Effect of Trust

• Choose a source
• Find all friends of the source who share a common
  neighbor with the source
• ? is the absolute difference between the sources
  rating and neighbors rating
• Categorize by sources rating of neighbor

N
7
9
?1
Source
Sink
6
67
(No Transcript)
68
(No Transcript)
69
Average ? for Trust values and Path Lengths
70
(No Transcript)
71
FilmTrust Recommendations

• Weight ratings by trust value
• Rating r from node s to movie m
• E.G.
• Alice rates Jaws at 3 stars
• Bob rates Jaws at 1 star
• Chuck trusts Alice at level 8
• Chuck trusts Bob at level 2
• The recommended rating for Chuck
• 83 21 / (82) 26/10 2.6 Stars

72
Complexity and Accuracy

• Complexity O(VE) - basically O(E) for our
  networks, where EV
• Comparison to other algorithms
• Beth-Borcherding-Klein (BBK) 1994