27
Propagation of Propagation of Trust and Distrust Trust and Distrust R. Guha, R. Kumar, P. Raghavan and A. Tomkins New York, 2004 Antti Sorjamaa Antti Sorjamaa Time Series Prediction and Time Series Prediction and ChemoInformatics Group ChemoInformatics Group Adaptive Informatics Research Centre Adaptive Informatics Research Centre Helsinki University of Technology Helsinki University of Technology

Time Series Prediction and ChemoInformatics Group Adaptive Informatics Research Centre

  • Upload
    amalie

  • View
    41

  • Download
    0

Embed Size (px)

DESCRIPTION

Propagation of Trust and Distrust R. Guha , R. Kumar, P. Raghavan and A. Tomkins New York, 2004 Antti Sorjamaa. Time Series Prediction and ChemoInformatics Group Adaptive Informatics Research Centre Helsinki University of Technology. Outline. Background Algorithms Atomic propagation - PowerPoint PPT Presentation

Citation preview

Page 1: Time Series Prediction and  ChemoInformatics  Group Adaptive Informatics Research Centre

Propagation of Propagation of Trust and DistrustTrust and Distrust

R. Guha, R. Kumar, P. Raghavan and A. Tomkins

New York, 2004

Antti SorjamaaAntti Sorjamaa

Time Series Prediction and ChemoInformatics Time Series Prediction and ChemoInformatics GroupGroupAdaptive Informatics Research CentreAdaptive Informatics Research CentreHelsinki University of TechnologyHelsinki University of Technology

Page 2: Time Series Prediction and  ChemoInformatics  Group Adaptive Informatics Research Centre

OutlineOutline BackgroundBackground AlgorithmsAlgorithms

– Atomic propagationAtomic propagation– Propagation methodsPropagation methods– RoundingRounding– Transitivity issuesTransitivity issues

Experimental resultsExperimental results

Antti Sorjamaa - TSPCi - AIRC - HUTAntti Sorjamaa - TSPCi - AIRC - HUT 22/27/27

Page 3: Time Series Prediction and  ChemoInformatics  Group Adaptive Informatics Research Centre

BackgroundBackground Trust Trust Valid information Valid information Distrust Distrust Disinformation or no Disinformation or no

informationinformation Increasing flow of information makes it Increasing flow of information makes it

necessary to validate the informationnecessary to validate the information Opinion of a single user is always Opinion of a single user is always

right, but is it trustworthy?right, but is it trustworthy? eBay, Epinions, huuto.net, etc.eBay, Epinions, huuto.net, etc.

Antti Sorjamaa - TSPCi - AIRC - HUTAntti Sorjamaa - TSPCi - AIRC - HUT 33/27/27

Page 4: Time Series Prediction and  ChemoInformatics  Group Adaptive Informatics Research Centre

Background (2)Background (2) Problems of disinformationProblems of disinformation

– Stock manipulation by bogus Stock manipulation by bogus postings on investment boardspostings on investment boards

– Marketers posing as customersMarketers posing as customers– Search engine link spammingSearch engine link spamming– Online relationships, datingOnline relationships, dating

Web of TrustWeb of Trust– Does not apply directly to DistrustDoes not apply directly to Distrust

Antti Sorjamaa - TSPCi - AIRC - HUTAntti Sorjamaa - TSPCi - AIRC - HUT 44/27/27

Page 5: Time Series Prediction and  ChemoInformatics  Group Adaptive Informatics Research Centre

Origin of TrustOrigin of Trust Collect opinions of several usersCollect opinions of several users

– One user can use several accountsOne user can use several accounts– Group of users agree and perform spoofingGroup of users agree and perform spoofing

IRL Trust is built and maintained over IRL Trust is built and maintained over long time periodslong time periods– Trust relationshipsTrust relationships– Good basis for the propagation of TrustGood basis for the propagation of Trust– More individual view of the Web of TrustMore individual view of the Web of Trust

Antti Sorjamaa - TSPCi - AIRC - HUTAntti Sorjamaa - TSPCi - AIRC - HUT 55/27/27

Page 6: Time Series Prediction and  ChemoInformatics  Group Adaptive Informatics Research Centre

Problems of DistrustProblems of Distrust ””Negative Trust”Negative Trust” Shifting the already existing Trust Shifting the already existing Trust

scores distort the resultsscores distort the results Algorithmic complicationsAlgorithmic complications

– Negative eigenvaluesNegative eigenvalues– Negative probabilitiesNegative probabilities

What is Distrust propagation?What is Distrust propagation?

Antti Sorjamaa - TSPCi - AIRC - HUTAntti Sorjamaa - TSPCi - AIRC - HUT 66/27/27

Page 7: Time Series Prediction and  ChemoInformatics  Group Adaptive Informatics Research Centre

NotationNotation

Antti Sorjamaa - TSPCi - AIRC - HUTAntti Sorjamaa - TSPCi - AIRC - HUT

nn = number of users = number of usersT = Trust matrix, T = Trust matrix, nn x x nn, values from 0 to 1, values from 0 to 1D = Distrust matrix, as Trust matrixD = Distrust matrix, as Trust matrix

B = Beliefs matrix, typically T or T - DB = Beliefs matrix, typically T or T - DCCB,B,αα = Combined atomic propagation matrix= Combined atomic propagation matrixPP((k k ) ) = Propagation matrix using = Propagation matrix using kk steps stepsF = Final beliefsF = Final beliefs

77/27/27

Page 8: Time Series Prediction and  ChemoInformatics  Group Adaptive Informatics Research Centre

Atomic PropagationAtomic Propagation Basic propagation techniques, ”atoms”Basic propagation techniques, ”atoms” Matrix operations, multiplying B with Matrix operations, multiplying B with

appropriate matrix operatorappropriate matrix operator 4 basic atoms4 basic atoms

– Direct PropagationDirect Propagation– Co-citationCo-citation– Transpose TrustTranspose Trust– Trust couplingTrust coupling

Antti Sorjamaa - TSPCi - AIRC - HUTAntti Sorjamaa - TSPCi - AIRC - HUT

1 2 3

4

1 2

35

Page 9: Time Series Prediction and  ChemoInformatics  Group Adaptive Informatics Research Centre

Atomic Propagation (2)Atomic Propagation (2) Direct Propagation, BDirect Propagation, B Co-citation, BCo-citation, BTTBB Transpose Trust, BTranspose Trust, BTT

Trust coupling, BBTrust coupling, BBTT

Antti Sorjamaa - TSPCi - AIRC - HUTAntti Sorjamaa - TSPCi - AIRC - HUT

1 2 3

0 1 00 0 0,70 0 0

0 1 00 0 0,70 0 0 =

0 0 0,70 0 00 0 0

99/27/27

Page 10: Time Series Prediction and  ChemoInformatics  Group Adaptive Informatics Research Centre

4

1

35Trust Coupling, BBTrust Coupling, BBTT

Antti Sorjamaa - TSPCi - AIRC - HUTAntti Sorjamaa - TSPCi - AIRC - HUT

0 0 0 1 00 0 0 0 00 0 0 1 00 0 0 0 00 0 1 0 0

=

0 0 0 0 00 0 0 0 00 0 0 0 00 0 0 0 00 0 0 1 0

0 0 0 1 00 0 0 0 00 0 0 1 00 0 0 0 00 0 1 0 0

=

0 0 0 0 00 0 0 0 00 0 0 0 00 0 0 0 00 0 0 1 0

0 0 0 0 00 0 0 0 00 0 0 0 11 0 1 0 00 0 0 0 0

0 0 0 0 00 0 0 0 00 0 0 0 00 0 0 0 01 0 1 0 0

1010/27/27

Page 11: Time Series Prediction and  ChemoInformatics  Group Adaptive Informatics Research Centre

Atomic Propagation (3)Atomic Propagation (3) Direct Propagation, BDirect Propagation, B Co-citation, BCo-citation, BTTBB Transpose Trust, BTranspose Trust, BTT

Trust coupling, BBTrust coupling, BBTT

Antti Sorjamaa - TSPCi - AIRC - HUTAntti Sorjamaa - TSPCi - AIRC - HUT 1111/27/27

Page 12: Time Series Prediction and  ChemoInformatics  Group Adaptive Informatics Research Centre

NotationNotation

Antti Sorjamaa - TSPCi - AIRC - HUTAntti Sorjamaa - TSPCi - AIRC - HUT

nn = number of users = number of usersT = Trust matrix, T = Trust matrix, nn x x nn, values from 0 to 1, values from 0 to 1D = Distrust matrix, as Trust matrixD = Distrust matrix, as Trust matrix

B = Beliefs matrix, typically T or T - DB = Beliefs matrix, typically T or T - DCCB,B,αα = Combined atomic propagation matrix= Combined atomic propagation matrixPP((k k ) ) = Propagation matrix using = Propagation matrix using kk steps stepsF = Final beliefsF = Final beliefs

1212/27/27

Page 13: Time Series Prediction and  ChemoInformatics  Group Adaptive Informatics Research Centre

Propagation MethodsPropagation Methods

Trust only: B = TTrust only: B = T

One-step Distrust: B = TOne-step Distrust: B = T

Propagated Distrust: B = T - DPropagated Distrust: B = T - D

Antti Sorjamaa - TSPCi - AIRC - HUTAntti Sorjamaa - TSPCi - AIRC - HUT 1313/27/27

Page 14: Time Series Prediction and  ChemoInformatics  Group Adaptive Informatics Research Centre

Iterative PropagationIterative Propagation Eigenvalue PropagationEigenvalue Propagation

Weighted Linear Combinations Weighted Linear Combinations (WLC)(WLC)

Antti Sorjamaa - TSPCi - AIRC - HUTAntti Sorjamaa - TSPCi - AIRC - HUT 1414/27/27

Page 15: Time Series Prediction and  ChemoInformatics  Group Adaptive Informatics Research Centre

RoundingRounding Binary decision of Trust (±1)Binary decision of Trust (±1)

– Global RoundingGlobal Rounding Order F according to Trust valuesOrder F according to Trust values Fraction of Trust in the whole FFraction of Trust in the whole F

– Local RoundingLocal Rounding Order F according to Trust valuesOrder F according to Trust values Fraction of Trust in the corrent row of FFraction of Trust in the corrent row of F

– Majority RoundingMajority Rounding Smallest local well-defined neighborhoodSmallest local well-defined neighborhood

Antti Sorjamaa - TSPCi - AIRC - HUTAntti Sorjamaa - TSPCi - AIRC - HUT 1515/27/27

Page 16: Time Series Prediction and  ChemoInformatics  Group Adaptive Informatics Research Centre

0 1 00 0 0,70 0 0

0 -1 00 0 -

0,70 0 0

TransitivityTransitivity Direct Propagation of TrustDirect Propagation of Trust

– If 1 trusts 2 and 2 trusts 3 then 1 trusts 3If 1 trusts 2 and 2 trusts 3 then 1 trusts 3 Does not apply to Distrust directlyDoes not apply to Distrust directly

””Enemy of my enemy is my friend”Enemy of my enemy is my friend” Multiplicative Trust PropagationMultiplicative Trust Propagation

Antti Sorjamaa - TSPCi - AIRC - HUTAntti Sorjamaa - TSPCi - AIRC - HUT

=0 -1 00 0 -

0,70 0 0

0 0 0,70 0 00 0 0

1616/27/27

Page 17: Time Series Prediction and  ChemoInformatics  Group Adaptive Informatics Research Centre

Transitivity (2)Transitivity (2) Multiplicative Distrust creates problemsMultiplicative Distrust creates problems

– As seen on previous slide:As seen on previous slide:””Enemy of my enemy is my friend”Enemy of my enemy is my friend”

– Directed cycle with negative values can Directed cycle with negative values can lead to the user to distrust himself!lead to the user to distrust himself!

Additive Distrust PropagationAdditive Distrust Propagation– Transform the basic propagation atomTransform the basic propagation atom– For example exp(bFor example exp(bijij) ) b bijij

Antti Sorjamaa - TSPCi - AIRC - HUTAntti Sorjamaa - TSPCi - AIRC - HUT 1717/27/27

Page 18: Time Series Prediction and  ChemoInformatics  Group Adaptive Informatics Research Centre

Experimental ResultsExperimental Results Epinions dataset from Epinions.comEpinions dataset from Epinions.com Users write and read reviews from Users write and read reviews from

many different topicsmany different topics Users can also rate the reviews and Users can also rate the reviews and

other reviewersother reviewers Review reading creates royaltiesReview reading creates royalties

– Distrust system to decrease abuseDistrust system to decrease abuse Similar systems: Amazon, SlashdotSimilar systems: Amazon, Slashdot

Antti Sorjamaa - TSPCi - AIRC - HUTAntti Sorjamaa - TSPCi - AIRC - HUT 1818/27/27

Page 19: Time Series Prediction and  ChemoInformatics  Group Adaptive Informatics Research Centre

Epinions DatasetEpinions Dataset 131 829 nodes with 841 372 edges131 829 nodes with 841 372 edges Edges either Trust (+1) or Distrust (-1)Edges either Trust (+1) or Distrust (-1) 85 percent of edges are Trust edges85 percent of edges are Trust edges Large Strongly Connected Component Large Strongly Connected Component

(SCC) of over 41 000 nodes(SCC) of over 41 000 nodes Bowtie: In almost 40 000 nodes and out Bowtie: In almost 40 000 nodes and out

more than 30 000 nodesmore than 30 000 nodes

Antti Sorjamaa - TSPCi - AIRC - HUTAntti Sorjamaa - TSPCi - AIRC - HUT 1919/27/27

Page 20: Time Series Prediction and  ChemoInformatics  Group Adaptive Informatics Research Centre

MethodologyMethodology Propagation of DistrustPropagation of Distrust

– Trust, One-step and Propagated DistrustTrust, One-step and Propagated Distrust Iteration Methods: EIG and WLCIteration Methods: EIG and WLC Rounding: Global, Local and MajorityRounding: Global, Local and Majority Atomic Propagations:Atomic Propagations:

– Direct, Co-sitation and all combinedDirect, Co-sitation and all combined 81 different combinations81 different combinations

Antti Sorjamaa - TSPCi - AIRC - HUTAntti Sorjamaa - TSPCi - AIRC - HUT 2020/27/27

Page 21: Time Series Prediction and  ChemoInformatics  Group Adaptive Informatics Research Centre

ExperimentsExperiments Each combination of methods get the Each combination of methods get the

same treatmentsame treatment Validation using Leave-One-Out (LOO)Validation using Leave-One-Out (LOO)

– Total of 3250 random edgesTotal of 3250 random edges prediction error prediction error εε

– About 500 Distrust and 500 Trust edgesAbout 500 Distrust and 500 Trust edges prediction error prediction error εεSS

Naive errors: Naive errors: εε = 0,15 and = 0,15 and εεSS = 0,5 = 0,5

Antti Sorjamaa - TSPCi - AIRC - HUTAntti Sorjamaa - TSPCi - AIRC - HUT 2121/27/27

Page 22: Time Series Prediction and  ChemoInformatics  Group Adaptive Informatics Research Centre

2222/27/27Antti Sorjamaa - TSPCi - AIRC - HUTAntti Sorjamaa - TSPCi - AIRC - HUT

ResultsResults

Page 23: Time Series Prediction and  ChemoInformatics  Group Adaptive Informatics Research Centre

Results (2)Results (2)

Antti Sorjamaa - TSPCi - AIRC - HUTAntti Sorjamaa - TSPCi - AIRC - HUT 2323/27/27

Page 24: Time Series Prediction and  ChemoInformatics  Group Adaptive Informatics Research Centre

Results (3)Results (3)

Antti Sorjamaa - TSPCi - AIRC - HUTAntti Sorjamaa - TSPCi - AIRC - HUT 2424/27/27

Page 25: Time Series Prediction and  ChemoInformatics  Group Adaptive Informatics Research Centre

Results (4)Results (4)

Antti Sorjamaa - TSPCi - AIRC - HUTAntti Sorjamaa - TSPCi - AIRC - HUT 2525/27/27

Page 26: Time Series Prediction and  ChemoInformatics  Group Adaptive Informatics Research Centre

ConclusionsConclusions Web of Trust and Distrust is important Web of Trust and Distrust is important

part of many e-commerce related sitespart of many e-commerce related sites– Distrust stabilizes the propagation of Distrust stabilizes the propagation of

opinions in the networkopinions in the network– Distrust decreases the effect of abuseDistrust decreases the effect of abuse

Rounding is surprisingly importantRounding is surprisingly important Small number of expressed Trust Small number of expressed Trust

scores lead to accurate predictionscores lead to accurate prediction

Antti Sorjamaa - TSPCi - AIRC - HUTAntti Sorjamaa - TSPCi - AIRC - HUT 2626/27/27