Reputation

1

• Reputation
• Prakash Kolan
• Liqin Zhang
• Venkatesh Kancherla

2
Introduction

• Internet
• No longer just a medium for non-commercial
  informal information exchange between scientists
  and universities13
• It has become a public network also used to
  support commercial transactions
• Unclear what will happen when this extremely open
  network is used in the new context of commerce
• Likely that the introduction of money will be the
  motivation for criminal activities previously
  considered uninteresting.

3
Introduction

• Expansion of the Internet
• People and services are called upon to interact
  with independent parties in application areas
  like e-commerce, knowledge sharing, game playing
  etc.
• Anyone is free to add what components (hardware
  and software) as he/she wishes
• No central authority keeps track of who is using
  it and how
• An electronic market with a centralized verifying
  authority that checks and certifies (human and
  electronic) participants would be a very non-open
  solution

4
Introduction

• Parties are autonomous and potentially subject to
  different administrative and legal domains
• Important that
• Decentralized and open mechanisms exist that
  allow participants in a market to know something
  about other participants
• Each participant should be able to identify
  trustworthy parties or correspondents with whom
  they should interact and untrustworthy
  correspondents with whom they should avoid
  interaction without having to rely on some
  external central authority

5
Need for Reputation

• We need Reputation because
• Internet is as an open system more like a big
  city than a small village13
• Possible to act in any possible way without
  anyone being able to stop it.
• Large amount of fraud and con men doing
  businesses and lots of harmful content floating
  out there
• In a big city you cant know who you are dealing
  with if you meet for the first time

6
Need for Reputation

• We need Reputation because
• How is it considered possible to negotiate,
  cooperate and perform online communication if
  there is no way of formally knowing the
  intentions of the other participants

7
Defining Reputation

• According to Oxford dictionary, reputation is the
  Common or general estimate of a person with
  respect to characters or other qualities5
• Reputation refers to a perception that an agent
  has of anothers intentions and norms15
• An entitys reputation is some notion or report
  of its propensity to fulfill the trust placed in
  it (during a particular situation) its
  reputation is created through feedback from
  individuals who have previously interacted with
  the entity16.

8
Defining Reputation

• Reputation, a distributed knowledge phenomenon,
  lives in time. When people interact with one
  another over time, the history of their past
  interactions informs others about their abilities
  and depositions17.
• Reputation systems are complex social systems
  that continually collect, aggregate, and
  distribute feedback about a person, an
  organization, a scholarly work, or some other
  entity, based on the assessments of others from
  their interactions or experiences with the
  entity17.

9
Reputation Systems

• Types of Reputation Systems - Dimensions of
  Classification
• Amount of effort required of users to generate
  reputations.
• Explicit action by the users, such as giving
  ratings and scores
• Users behavior, such as return rates
• Ease of understanding by the user
• Ease of implementation for developers
• Degree of personal relevance of ratings to users
• Personal relevance is the degree to which ratings
  take into consideration the users likes and
  dislikes or the extent to which recommendations
  are tailored to the individual user

10
Reputation Systems

• Reputation systems can be grouped according to
  the nature of information they give about the
  object of interest and how the rating is
  generated.
• Ranking Systems
• Rating Systems
• Collaborative Filtering Systems
• Peer Based Reputation systems
• Implicit Peer Based Reputation systems
• Explicit Peer Based Reputation systems

11
Reputation Systems

• Ranking Systems
• Use quantifiable measures of users behavior
  (implicit information) to generate a rating.
• Example ranking systems High score lists,
  information about length of membership, frequency
  of visits, replies etc.
• Easy to implement and interpret and are most
  suited for goal oriented activities
• These reputation systems typically only provide
  information about what kind of pattern users
  follow, and reveal little or no personally
  relevant information.

12
Reputation Systems

• Rating Systems
• Use explicit evaluations given by users.
• These evaluations are used to generate a weighted
  average for each object of interest.
• Ratings are global, meaning that all users
  looking at the same object of interest will see
  the same score.
• Provide more personally relevant information than
  ranking systems, they treat the population as a
  single homogenous group.

13
Reputation Systems

• Collaborative Rating Systems
• These systems weight explicit or implicit
  evaluations by how much the rater and the user
  have concurred on other items
• More sophisticated than rating systems, capturing
  significant amounts of personally relevant
  informationusers likes and dislikes.
• Most expensive to build, populate, maintain, as
  well as the most complicated for users to
  understand

14
Reputation Systems

• Peer Based Reputation Systems
• Based on peer recommendations like friends and
  family
• Peer-based recommendations (or social network
  based reputation systems), whether they are given
  explicitly or inferred through the observations
  of peer behavior, are a significant influence on
  everyday decision-making
• The social context provided by friend of a
  friend recommendations should be especially
  important in socially-oriented situations
• The more social the situation, the more important
  peer based information is.

15
Reputation Systems

• Implicit Peer Based Reputation Systems
• These systems track the behavior of a users
  friends, generating ratings from this data.
• Such systems observe what a users friends do
  (e.g., with whom they interact, what they look
  at, what they buy), and make recommendations
  accordingly.
• These types of systems is that they provide
  information that is very socially relevant and
  tailored to the individual.
• Potential drawbacks are the implementation costs,
  privacy concerns, and that such ratings might be
  difficult to understand for users

16
Reputation Systems

• Explicit Peer Based Reputation Systems
• These systems rely on the evaluations given by a
  users friends
• Users select a group of friends or trusted
  raters, and the evaluations made by this group
  are used to generate composite ratings.
• These system weights or filters ratings based on
  who we know and choose to trust.
• Ratings are highly relevant and tailored to the
  user.
• Drawbacks include implement costs and difficulty
  in understanding

17
Reputation Systems
18
Notions of Reputation
Reputation can be viewed as a global or
personalized quantity15
Reputation Typology
19
Notions of Reputation

• Individual Group reputation

Reputation is a function of the cumulative
ratings on users by others for a individual
A firms (group) reputation can be modeled as the
average of all its members individual reputation
20
Notions of Reputation

• Direct Indirect( Individual) reputation

Reputation estimates by an evaluator based on
direct experiences (seen or experienced by the
evaluating agent first hand)
Reputation estimates that are based on
second-hand evidence (such as by word-of-mouth).
21
Notions of Reputation

• Direct Reputation

Reputation based on actual encounter with the
reputed agent
Reputation based on evaluators rating for the
reputed agent
22
Notions of Reputation

• Indirect Reputation

Reputation based on the prior belief regarding
the reputed agent
Reputation for the reputed agent based on the
group he belongs to
Reputation garnered from different evaluating
agents for the reputed agent
23
Requirements
24
Challenges in Eliciting feedback

• The first is that people may not bother to
  provide feedback at all. For example, when a
  trade is completed successfully at eBay, there is
  little incentive to spend another few minutes
  filling out a form
• People could be paid for providing feedback
• Secondly,It is especially difficult to elicit
  negative feedback. For example, at eBay it is
  common practice to negotiate first before
  resorting to negative feedback. Therefore, only
  really bad performances are reported.

25
Challenges in Eliciting feedback

• The third difficulty is assuring honest reports.
• One party could blackmail anotherthat is,
  threaten to post negative feedback unrelated to
  actual performance.
• At the other extreme, in order to accumulate
  positive feedback a group of people might
  collaborate and rate each other positively,
  artificially inflating their reputations.

26
Challenges in Distributing feedback

• The first is name changes. At many sites, people
  choose a pseudonym when they register. If they
  register again, they can choose another
  pseudonym, effectively erasing prior feedback.
• Two methods to avoid Name Changes
• Game theoretic analysis
• Another alternative is to prevent name changes,
  either by using real names, or by preventing
  people from acquiring multiple pseudonyms, a
  technique called once-in-a-lifetime pseudonyms

27
Challenges in Distributing feedback

• A second difficulty in distributing feedback
  stems from lack of portability between systems.
• Amazon.com initially allowed users to import
  their ratings from eBay. eBay protested
  vigorously, claiming that their user ratings were
  proprietary. Ultimately Amazon discontinued its
  rating-import service.
• Efforts are underway to construct a more
  universal framework. For example,
  virtualfeedback.com provides a rating service for
  users across different systems, but it has yet to
  gain wide public acceptance.

28
Challenges in Distributing feedback

• Finally,There is also a potential difficulty in
  aggregating and displaying feedback so that it is
  truly useful in influencing future decisions
  about who to trust.
• eBay displays the net feedback (positives minus
  negatives). Other sites such as Amazon display an
  average.

29
Context and location awareness

• Another important consideration is the
  context and location awareness, as many of the
  applications are sensitive to the context or the
  location of the transactions.
• For example, the functionality of the transaction
  is an important context to be incorporated into
  the trust metric. Amazon.com may be trustworthy
  on selling books but not on providing medical
  devices.

30
Different methods

• Basic models
• Reputation models in peer-to-peer networks
• Reputation models in social networks

31
Rating systems

• Reputation is taken to be a function of the
  cumulative positive or negative rating for a
  seller or buyer
• Rating model
• Uniform context environment heard rating from
  one agent
• Multiple context environment from multiple
  agents
• Centrality-based rating based on in/out degree
  of a node
• Preference-based rating Consider the preferences
  of each member when selecting the reputable
  members
• Bayesian estimate rating to compute reputation
  with recommendation of different context

32
Basic models

• Computational model
• Based on how much deeds exchanged
• Collaborative model
• Based on recommendation from similar tasted people

33
Computational model2

• If Reputation increase, trust increase
• If trust increase, reciprocity increase
• If reciprocity increase, reputation increase

Reputation
Reciprocity mutual exchange of deeds
Net benefit
Reciprocity
Trust
34
A Collaborative reputation mechanism

• Collaborative filtering
• To detect patterns among opinions of different
  users
• Make recommendation based on rating of people
  with similar taste
• Fake rating
• 1. Rate more than once
• 2. Fake identity
• Solve rating from people with high reputation in
  network weighted more

35
Reputation model in peer-to-peer11

• P2P network
• peers cooperate to perform a critical function
  in a decentralized manner
• Peers are both consumers and providers of
  resources
• Peers can access each other directly
• Allow peers to represent and update their trust
  in other peers in open networks for sharing files

36
Models in peer-to-peer networks

• Based on recommendation from other peers
• Combine with Bayesian network
• Based on global trust value

37
Method 1 Reputation based on recommendation 11

38

• Recomendation from different kind of peers
• Different weight
• Update references weight
• Final reputation and trust is computed based on
  Bayesian network
• Solve reputation on different aspects of a peer

39
Method2 based on global trust value---Eigen
Trust Algorithm12

• Decreases the number of downloads of
  unauthenticated files in a peer-to-peer file
  sharing network by assigning a unique global
  trust value
• A distributed and secure method to compute global
  trust values based on power iteration
• Peers use these global trust values to choose the
  peers from whom they download and share files

40
Reputation Peer to Peer N/w

• Limited Reputation Sharing in P2P Systems14
• Techniques based on collecting reputation
  information which uses only limited or no
  information sharing between nodes.
• Effect of limited reputation information sharing
  in a peer-to-peer system.
• Efficiency
• Load distribution and balancing
• Message traffic

41
Reputation models in Social networks310

• Social network
• a representation of the relationships existing
  within a community
• Each node provide both services and referrals for
  services to each other

42
Importance of the nodes

• Proposal 1 all nodes are equal important
• Proposal 2 some nodes are important than others
• Referrals from A, B, C,D,E is more important than
  those nodes in only local network pivot
• You may trust the referral from a friend of you
  than strangers
• You may also need consider the your preference
  regarding to referral

43
Models in social network

• Reputation extracting model
• Ranking the reputation for each node in network
  based on their location
• Social ReGreT model
• Based on information collected from three
  dimension

44
Reputation models in Social networks

• Extracting Reputation in Multi agent systems8
• Feedback after interaction between agents
• Also consider the position of an agent in social
  network
• Node ranking creating a ranking of reputation
  ratings of community members
• Based on the in-degree and out-degree of a node
  (like Pagerank)

45
Reputation models in Social Networks

• Social ReGreT5
• Analysis social relation
• To identify valuable features in e-commerce
• Aimed to solve the problem of referrers false,
  biased or incomplete information
• Based on three dimensions of reputation
• If use only interaction inf. --- individual
  dimension(single)
• If also use inf. from others --- social dimension
  (multiple)
• Three dimension
• Witness reputation from pivot agents
• Neighborhood reputation
• System reputation default reputation value based
  on the role played by the target agent

46
Metrics

• The algorithm used to calculate an agents
  reputation is the metric of the reputation
  system.
• The strength of a metric is measured by its
  resistance against different threat models, i.e,
  different types of hostile agents.

47
Formal Model

• The model provides an abstract view of a
  reputation system that allows the comparison of
  the core metrics of different reputation systems.
• According to definition of reputation a
  transaction between two peers is the basis of a
  rating. An agent cannot rate another one without
  having had a transaction with him.

48
Formal Model

• A is the set of agents.
• C is the context of a transaction.
• set C T V
• where T 0, 1, . . . , tnow is the set of
  times and V is the set of transaction values
• E is the set of all encounters between different
  agents that have happened until now.

49
Formal Model

• An encounter contains information about the
  participating peers and the context
• A rating is a mapping between a target agent a
  belongs to A and an encounter e belongs to E
  to the set of all possible ratings Q
• In the simple case Q is a small set of possible
  values Qebay -1, 0, 1

50
Formal Model

• Ea represents the subset of all encounters in
  which a has participated and received a rating
• All encounters between a and b with a valid
  rating for a are
• subset of all most recent encounters between a
  and other agents.

51
Formal Model

• The reputation of an agent a belongs to A is
  defined by the function r A T -gtR.
• In short r(a)r(a,tnow)
• A complete Metric M is defined as
• M(p,r,Q,R,r0)

52
METRICS IN REPUTATION SYSTEM

• Accumulative Systems
• Average Systems
• Blurred Systems
• OnlyLast Systems
• EigenTrust System

53
Accumulative Systems

• If a system accumulates all given ratings to get
  the overall reputation of an agent we call it an
  accumulative system.
• Example Ebay system
• Possible ratings are p A E gt -1, 0, 1.
• The basic idea of these metrics is, that the more
  often an agent behaves in a good way the more
  sure can the others be, that this agent is an
  honest one.

54
Accumulative Systems

• The reputation of an agent a belongs to A
• computes with
• 
  (ebay)
• No transaction values and multiple ratings
• The reputation in value system is given by

55
Average Systems

• This kind of reputation system computes the
  reputation for an agent as the average of all
  ratings the agent has received
• The idea of this metric is, that agents behave
  the same way most of their lifetime. Unusual
  ratings have only little weight in the
  computation of the final reputation
• The simulated systems use
• p A E -gt -1, 0, 1

56
Average system

• The reputation of an agent a belongs to A in
  the Average-system without considering multiple
  ratings and transaction values is
• In average-value system

57
Blurred Systems

• These reputation systems compute a weighted sum
  of all ratings.
• The older a rating is, the less it influences the
  current reputation
• Possible ratings are p A E -gt -1, 0, 1

58
Blurred System

• The reputation of an agent a belongs to A
  without considering transaction values is
• With consideration of transaction values

59
OnlyLast System

• This system considers the most recent rating of
  an agent
• Ratings are p A E -gt -1, 0, 1
• Here we expect an agent to behave like he did
  last time, no matter what he did before.

60
OnlyLast System

• Without considering transaction values in the
  OnlyLast system the reputation of an agent a
  belongs to A is
• With consideration of the transaction value in
  the OnlyLastValue system the reputation of an
  agent a belongs to A is

61
EigenTrust System

• In this metric the computed reputation depends on
  the ratings, the reputation of the raters, the
  transaction context (e.g. transaction value), and
  some community properties
• Ratings are p A E -gt -1,1
• First we have to build a reputation matrix M,
  where
• (mij) contains the standardized sum of ratings
  from Agent i for Agent j

62
EigenTrust System
63
(No Transcript)
64
New Metric

• We can combine different metrics to compensate
  for the individual weaknesses.
• Both Average and OnlyLast systems can be
  understood as summing up the previous ratings of
  an agent using different weights.
• The Blurred-system is somewhere in between, but
  could not handle the disturbing agents.

65
New Metric

• Thus we can interpolate between the Average and
  the OnlyLast-system by weighting the ratings not
  linear, like we did in the Blurred-system, but
  quadratic, so that the recent ratings have more
• influence on the reputation.
• The resulting metric M (p, r) is
• p A E -gt -1, 0, 1

66
New Metric

• We call this metric a BlurredSquared System
• This system is invulnerable to disturbing, evil,
  and selfish agents. It resists malicious agent up
  to an amount of 60.

67
Conclusions

• Reputation is very important in electronic
  communities
• Reputation can have different notation such as
  general estimate a person, perception that an
  agent has of anothers intentions and norms
• Reputation systems can be grouped according to
  the nature of information they give about the
  object of interest and how the rating is
  generated, 4 reputation systems are discussed

68
Conclusions

• Reputation can be classified to individual and
  group reputation, individual reputation can be
  further classified
• The challenge for reputation includes less
  feedback, negative feedback, un-honesty feedback
  (change name), context and location awareness
• An agent can be honesty, malicious, evil, selfish
• Discussed 7 metrics with benchmarks

69
Conclusions Comparison methods

• Basic models
• Computation model
• based on how much deeds exchanged
• Can be used in P2P and Social network
• Doesnt consider references/recommendation,
  weight of deeds
• Collaborative model
• Based on the recommendation from similar tasted
  people
• Recommendation is weighted based on referrers
  reputation avoid fake recommendation
• Doesnt consider the location of referrer

70
Conclusions Comparison methods

• In P2P network,
• Bayesian network model
• Based on information collected from friends
• Peers share recommendations
• It allows to develop different trust regarding to
  different aspects of the peers capability
• Overall trust need combine all aspect
• Doesnt consider location

71
Conclusions Comparison methods

• In social network
• Can consider the position of an agent, Pivot
  agents are more important than other agents
• NodeRanking
• Ranking the reputation in social network based on
  position
• Used to find the pivot
• Social ReGreT model
• Consider three dimension
• Witness pivot node
• Neighborhood recommendation
• System value

72
Conclusions

• The reputation computation need consider
  recommendation of friends, the position of the
  referrer, weight for referrer
• friends may refer to its neighborhood, or the
  group of people who has the similar taste, or
  people you trust
• Weight for referrer can avoid fake recommendation
• No models consider all of the factors

73
References

• 1. Computational Models of Trust and
  Reputation Agents, Evolutionary Games, and
  Social Networks, www.cdm.csail.mit.edu/ftp/lmui/
  computational20models20of20trust20and20reputa
  tion.pdf
• 2. A computation model of Trust and
  Reputation, http//csdl2.computer.org/comp/proceed
  ings/hicss/2002/1435/07/14350188.pdf
• 3. Trust and Reputation Management in a
  Small-World Network, ICMAS Proceedings of the
  Fourth International Conference on MultiAgent
  Systems (ICMAS-2000), 2000
• 4. How Social Structure Improves Distributed
  Reputation Systems, http//www.ipd.uka.de/nimis/p
  ublications/ap2pc04.pdf
• 5. Social ReGreT, a reputation model based on
  social relations , ACM SIGecom Exchanges Volume 3
  ,  Issue 1   Winter, 2002,Pages 44 56
• 6. Detecting deception in reputation
  management, Proceedings of the second
  international joint conference on Autonomous
  agents and multiagent systems , 2003

74
References

• 7. Finding others online reputation systems
  for social online spaces, Proceedings of the
  SIGCHI conference on Human factors in computing
  systems Changing our world, changing ourselves,
  2002, Pages 447 - 454  
• 8. J. Pujol and R. Sanguesa and J. Delgado,
  Extracting reputation in multi-agent systems by
  means of social network topology, In
  Proceedings of First International Joint pages
  467--474, 2002
• 9. J. Sabater and C. Sierra,Reputation and
  social network analysis in multi-agent systems,
  Proceedings of the first international joint
  conference on Autonomous agents and multiagent
  systems P475 482,2002
• 10. Trust evaluation through relationship
  analysis, Proceedings of the fourth international
  joint conference on Autonomous agents and
  multiagent systems,P1005 1011, 2005
• 11 Trust and Reputation model in peer-to-peer
  networks, www.cs.usask.ca/grads/
  yaw181/publications/120_wang_y.pdf

75
References

• 12 S. D. Kamvar, M. T. Schlosser, and H.
  Garcia-Molina. The Eigen Trust algorithm for
  reputation management in p2p networks. In
  Proceedings of the Twelfth International World
  Wide Web Conference, 2003.
• 13 Lars Rasmusson and Sverker Jansson,
  Simulated social control. for secure internet
  commerce, in New Security Paradigms 96.
  September 1996
• 14 S. Marti, H. Garcial-Molina, Limited
  Reputation Sharing in P2P Systems, ACM Conference
  on Electronic Commerce (EC'04)
• 15 Lik Mui, Computational Models of Trust and
  Reputation Agents, Evolutionary Games, and
  Social Networks, Ph. D Dissertation,
  Massachusetts Institute of Technology
• 16 Goecks, J. and Mynatt E.D. (2002). Enabling
  privacy management in ubiquitous computing
  environments through trust and reputation
  systems. Workshop on Privacy in Digital
  Environments Empowering Users. Proceedings of
  CSCW 2002

76
References

• 17 G.L. Rein, Reputation Information Systems A
  Reference Model, Proceedings of the 38th Hawaii
  International Conference on System Sciences - 2005