TPS
Trust and Provenance in
Sweto
Meenakshi Nagarajan
Willie Milnor
Nicole Oldham
Introduction
Nature of the Semantic Web
Machine understandable information
Open, distributed, low barriers with publication
New techniques to validate information
Provenance is key to establishing trust in
the information
Not adequate to associate trust in the
content of the source
Unreasonable to know trust in every
statement by verifying provenance and
source
Option1: Associate a trust value with every
source
CNN = 0.9
Counter-Intuitive to how we process
information
Statement about “War in Iraq” and “The Iraqi
People’s leader” made by CNN and Iraq
Daily.
Option 2: May be, a trust value for every
source for every domain under
consideration
Infinite domains and sources – not scalable
Option 3:
Possibilityof finite users ascertaining their
confidence in some statements
Trust anyone has on a statement as a
function of their trust on the user who placed
a confidence on this statement
Very close to humans analyze content to
ascertain credibility
Recommendation systems, e-Bay etc
TPS
Trusta member of a network can associate
with a statement on the Semantic Web is
proportional to the belief asserted on the
statement by some user (also in the network)
and the trust the member has on this user.
statement
have
beliefs in
User trusts Belief in statement
trusts
Ux
trusts trusts
trusts
trusts trusts
Based on this..
We identified the requirement of two
models
Provenance model (essentially Sweto itself)
Provenance information of statements
Trust model
Trust between users who placed a confidence
value in a statement in Sweto
Related Work
Knowledge management to determine the
validity and origin of information on the
web
http://www.eil.utoronto.ca/km/papers/fox-
kp1.pdf
Proof-like support system for explaining
provenance information
http://www.ksl.stanford.edu/people/pp/pap
ers/PinheirodaSilva_DEBULL_2003.pdf
Role of trust in ascertaining credibility of
information – Web of trust
http://www.cs.washington.edu/homes/pe
drod/papers/iswc03.pdf
A framework for trust propagation using
notions of trust and distrust in a web of
trust – e-commerce systems
http://tap.stanford.edu/trust04.pdf
Issues related to using trust in web
based social networks, specifically in
building and maintaining a trust network
on the web
http://trust.mindswap.org/
Combining trust and provenance
http://ebiquity.umbc.edu/v2.1/_file_direct
ory_/resources/58.pdf
The Models ..
Provenance Model – enhancing Sweto
Captures
Provenance information of statements in Sweto
Confidence / truth value of a statement
User who placed that confidence / truth value
The Models ..
Trust Model WOT
Captures
Trust between users, where a user E users who
entered a confidence / truth value in a statement
When a user enters a confidence / truth value into
the provenance model, he is
Added to the provenance model
Optionally, he could add himself to the WOT if he wishes
to place trust values in other users
Placing trust in other users of the WOT
intuitively,
user1 verifies the confidence value
placed by userx in the statement
Depending on the confidence values, user1
establishes trust in userx
A BIG ASSUMPTION
ALL USERS ARE BASICALLY TRUSTWORTHY AS FAR AS GOING THROUGH
THE PROCESS OF ENTERING TRUTH AND TRUST VALUES
Unique features and contribution
Features
Source and domain consideration. No single source,
single trust value concept
Personalized trust metrics for every user in the
system – respecting the subjective nature of trust
Adaptive model
Ability to change trust in users and/or truth values on
statements
Immediately reflects on results obtained
Aggregation in TPS
Primary Question we are trying to answer
How much can I trust an association I get
from Sweto ?
Can also answer
How much do I trust user x ? (directly or
through propagation of trust / distrust)
Web Of Trust
A directed Graph of users of the system with edge
weights as the trust values between them.
Every user who places a truth value in an assertion is
represented as a node in this graph.
B 0
0.7
E
A 0.7
0.2
0.8 0.3
1.0 D
0.4
C 0.6 F
Representation of Trust in the WOT
A matrix that contains the uA uB uC uD uE uF
actual trust values that each of tA 1.0 .7 1.0
the n users placed in any of
tB 1.0 0
the other users is maintained.
tC 1.0 .6
ti is the row representing the .2 .7 1.0 .4
tD
trust that user i has for each of
the other users. User i can tE .8 1.0
B 0
specify trust tik for any user k. 0.7
tF .3
E 1.0
A 0.7
If user i does not trust user k 0.2
then tik = 0. tik ≠ tki. 0.8 0.3
1.0 D
0.4
C 0.6 F
Propagation of Trust in the WOT
The trust will then be propagated throughout the WOT to obtain a matrix
that contains trust values for all users.
The trust value associated with each path is calculated by applying a
concatenation function to multiply the trusts along the path. For example,
tik * tkj is the amount that user i trusts user j via k.
ABED = 0 Aggregate Maximum for tAD is .6
A C D = .6
uA uB uC uD uE uF
The trust value tik will be
tA 1.0 .7 1.0 .6 .072 .24
recalculated as the trust
values change for any of tB 0 1.0 0 0 0 0
the users. tC .12 .42 1.0 .6 .072 .24
tD .2 .7 .2 1.0 .12 .4
tE .16 .55 .16 .8 1.0 .32
tF .048 .168 .048 .24 .3 1.0
Trust in a semantic association
Trust on a statement function of truth
value on the statement and trust on user
who placed this truth value
Extending this to a semantic association –
function of trusts on individual statements
Trust in a semantic association
Calculating trust in individual statements
Calculating trust in the association
User X
Calculating trust in a statement S
More than one user can place a truth value on
a statement
Trust in S = truth value placed on S by user
that user X trusts the most
Calculating trust in a semantic association
Only as strong as its weakest link.
The value of its least trustworthy component.
(statement)
TIPS Architecture
Web Interface
Trust ranking module Query
processor
(SemDis)
Trust
aggregator
Beliefs SWETO
WOT
Schema
WOT Beliefs
trusts truth_
user user value
with_probability
to_degree
believed_by
trust_ stmt user
value
Test Set
Small/manageable set of SWETO
instances
Synthetically generated 15 WOT users
Added corresponding nodes to the graph
Generated synthetic trust relationships
Random values between 0 and 1
Synthetically generated statements of truth
Random values between 0 and 1
Test Cases
1. A user requests both unranked and then ranked
results for the same query.
1. Unranked results appear in order found.
2. A user adds an explicit truth value to a statement in an
association.
1. All corresponding associations are affected
2. Some may be now have different ranks
3. A users changes/states and explicit trust in a believer
of a statement.
1. Corresponding associations are affected
2. Some now have different ranks
References
http://lsdis.cs.uga.edu/library/download/SAA+2004-PISTA.pdf
http://ebiquity.umbc.edu/v2.1/_file_directory_/resources/58.pdf
http://www.eil.utoronto.ca/km/papers/fox-kp1.pdf
http://www.ksl.stanford.edu/people/pp/papers/PinheirodaSilva_DEBULL_2003.pdf
http://www.cs.washington.edu/homes/pedrod/papers/iswc03.pdf
http://tap.stanford.edu/trust04.pdf
http://trust.mindswap.org/
http://lsdis.cs.uga.edu/projects/SemDis/Sweto/sweto.pdf
http://lsdis.cs.uga.edu/projects/SemDis/
http://lsdis.cs.uga.edu/lib/download/AS03-WWW.pdf
http://lsdis.cs.uga.edu/library/download/iswcRanking2004.pdf
http://tap.stanford.edu/trust04.pdf
http://www.cs.cornell.edu/home/kleinber/auth.pdf
http://www.semagix.com/
http://moloko.itc.it/paoloblog/papers/trust2004.pdf