270 likes | 381 Views
Reputation Management Survey. - Vinay. Introduction. Electronic markets, Distributed peer-to-peer applications - other forms of online collaboration All based on mutual trust, which enables transacting peers to overcome the uncertainty and risk inherent in the environment.
E N D
Reputation Management Survey - Vinay
Introduction • Electronic markets, • Distributed peer-to-peer applications • - other forms of online collaboration • All based on mutual trust, which enables transacting peers • to overcome the uncertainty and risk inherent in the • environment. Where does this Trust Come From ????
Reputation Systems • - Reputation systems provide essential input for computational trust • as predictions on future behavior based on a peer’s past • actions. • Information about these actions can also be received from other • members of a reputation network who have transacted with the peer But the credibility of this kind of Information must be critically assessed as it is third party information. - The Main goal of these systems is predicting a peer’s future actions, given knowledge about its past behavior - From Where do they get this Knowledge from ????
Reputation Systems • Ideally these kind of information comes from first-hand • transactions with that particular peer • But Interacting with every peer becomes costly • To overcome or lessen the cost Peers share their experiences • through the reputation systems • This makes it possible for the entire community to detect and • isolate misbehaving peers effectively. • - We refer to this shared information as recommendations.
Credibility Assessment • Reputation systems have been widely used for various • applications • Credibility Assessment of the provided reputation • information is becoming increasingly important where more • resources and business value depend on making correct trust • decisions • This Assessment is necessary because • A peer can provide inaccurate information. • Two peers can just come to an agreement and claim positive • experiences of each other to increase their reputation • A peer can defame another peer by providing bad experience to • to increase its own fame
- All these kinds of misbehavior in a reputation system • makes it important to separate accurate information from inaccurate • A thorough credibility analysis increases the system’s resistance • against malicious peers Here we are going to analyze the Maturity of current reputation systems . We are going to compare 11 reputation systems in terms of • The creation and content of a recommendation • The selection and use of recommenders • The interpretation and reasoning applied • to the gathered information
The terms which are mentioned before are the credibility aspects • The Creation and Content of Recommendation : • It determines whether the information in a recommendation provides sufficient basis for credibility analysis • It involves combining local experience information into a standard • Form to present to another peer. This can include both positive and negative • Experiences . Recommendation can include time of transaction to separate • new information from old. • If the recommendation passes through mediator then additional • information is needed to ensure transparency.
2. Selection and Use of Recommenders A node which is gathering recommendation information selects its recommenders from a set of possible nodes or it can simply allow and expect any knowledgeable node provide. The node which is gathering can be trustor ie a centralized entity collecting all recommendations. The main criterion to consider for selection is the credibility of the recommender 3. Interpretation and reasoning The acquired recommendations are combined with local experience ( trustor’s local experience ) and aggregated into suitable format. Finally the system may analyze the overall credibility of the Information.
Earlier we mentioned that we are going compare Reputation Systems They are • eBay • Unitec • FuzzyTrust • REGRET • NICE • Managing the Dynamic Nature of trust (MDNT) • PeerTrust • Managing Trust • Maximum Likelihood Estimation of Peer’s Performance • Eigen Trust • Travos These Reputation systems represent a wide range of applications with different requirements
eBay This reputation system stores users ratings linked to their profiles and elated transactions but leaves credibility analysis to human users Unitec This also leaves the credibility analysis to humans but apart from that it also performs a automated credibility analysis as well. FuzztTrust and REGRET Both designed for same purpose multi agent market place. But have different approaches.
FuzzyTrust In This reputation system, Local trust scores are generated and then added to global reputation values REGRET Viewpoints are applied as needed to enter a local reputation view , based on social relations between peers. NICE For Cooperative applications , Trustors are given signed receipts of successful transactions ie cookies as sign of trust
MDNT In This reputation estimate involves predicting trustee’s behavior based on experience from a specific time period MLE This uses probabilistic approach and considers the probability of recommenders to provide incorrect information PeerTrust It considers transaction and community contexts when estimating reputation
Managing Trust This reputation system considers only negative experiences and allows recommenders to remain anonymous EigenTrust This is file sharing system , It relies specifically on a global , shared view of reputation Travos The Reputation information is used to choose the most trustworthy partner. It bases reputation expression on probability distributions.
Recommendation Creation and Content in these reputation Systems When Creating a Recommendation , Recommender analyzes the local experience and also takes into consider the negative and Positive experience Managing trust In Managing Trust , it considers only negative experience And is the only system which allows sources of complaints anonymous. It also uses digital signature NICE and Unitec They use aggregated opinions but they do not specify the aggregation method and simply assume a policy is in place for determining it . This policy can be shared among all peers or locally defined.
REGRET An opinion is calculated as the weighted average of single experience ratings with more weight to given to newer experiences. Travos and EigenTrust They keep counters of positive and negative experiences and use them to provide aggregated opinion. Travos simply presents both counters. Eigen Trust calulates their difference and normalizes the value between the range [0,1]
MDNT In this system , opinion presented is calculated by CCCI metric CCCI Stands for Correlation Commit Clear Influence For Commit , we define 7 levels of criteria ie -1 to 5 None or ignore Nothing is delivered and so on to Fully Delivered all the Commitment For Clear Criteria , we again define 7 levels of criteria -1 to 5 None or ignore Not clear, barely clear, ….. Very clear
For Influence Criterion also 7 levels None or ignore Unimportant , Barely , Partially , largely , important, very imp. Then the Correlation is caluculated Correlation is calculated as Corr N = Σ(Commit criterion c * Clear criterion c * Inf criterion c ) c=1 This gives us how much the Trusted agent has delivered his commitment.
eBay The reputation System resides on a centralized server. It shows all ratings as part of user profiles and user can add their comment on it. MLE This is same as eBay but suggests to have time stamps to be included for the recommendation part as a part of the key for storing it. It also includes recommenders digital signature to guarantee its integrity.
Selection and Use of Recommenders in These Reputation Systems Out of the 11 reputation systems Five of them uses recommendations from all peers for selection . They are eBay Peertrust ManagingTrust EigenTrust Travos Two of them Limit the number of recommenders REGRET which groups peers according to their social context such as their relationship with trustee and then chooses the most representative member of each group
FuzzyTrust Recommenders are chosen through global weighting based on their number of performed transactions and local trust score. The weight is then compared to a threshold based on peer’s role , A Superpeer with high threshold than regular peer for load balancing. MDNT In this recommenders are selected based on their credibility. Which lowers the likelihood of new recommenders to get their voice Unitec It may have bias in selection as it is left to the user to decide.
MLE Reputation estimation is based only on existing recommendations but selection method for this s not defined. The most common way is to leave it to the trustor for slection. NICE Here the trustee tries to locate chains of first hand recommendations from trustor to itself
Reasoning and Interpretation in Reputation Systems Once the Reputation Information is collected , it must be aggregated into a suitable format then we examine how recommendations are aggregated and interpret them. The Final Interpretation of the result is most commonly threshold based. If the trustee’s reputation value is high enough , the trustor will decide to transact with it. This Approach is taken by NICE, MDNT, Managing Trust and MLE
FuzzyTrust, REGRET and PeerTrust do not specify how the final measure is interpreted but both threshold and rank based approaches are possible. Travos Explicitly uses rank based approach where a trustor orders a group of Potential partners based on their reputation for selection purpose. Generally Rank Based selection is only usable when several potential partners can provide a similar service and therefore interchangeable. eBay and Unitec do not perform interpretation at all as they present a report to a human user
Evaluating the result Evaluation is done rarely . Ad not all the systems do evaluation REGRET calculates a Reliability value for each type of reputation based on variety of factors such as available information, social relationships . eBay and Unitec provide reports for which given metadata evaluation is possible
Conclusion There are two points of focus for designing a reputation system to perform well in its field. Social Requirements & Scalability In this survey we have focused on analyzing the Social Requirements but there was not sufficient information available on Scalability of the analyzed systems. Scalability requires addressing three load challenges. Load placed on trustor, high reputation nodes and network through recommendation transfers. Allowing different levels of analysis based on situation is a solution for the above mentioned loads in Scalability
To make the Reputation Systems more mature • The Reputation Information should be standardized to achieve interoperability • Experience based reputation information should be based on commonly acceptable framework.