Computations In Social Networks Sajid S Shaikh Advisor: Javed I Khan.
-
date post
19-Dec-2015 -
Category
Documents
-
view
216 -
download
1
Transcript of Computations In Social Networks Sajid S Shaikh Advisor: Javed I Khan.
Computations In Social Networks
Sajid S Shaikh
Advisor: Javed I Khan
Structure Introduction
Related Work
Reputation Assessment System
Reputation Quantification System (RQS)
Reputation Reasoning System (RRS)
Reputation Based Social Computing
Social Network Based Computations
Classification
Examples
Conclusion
Introduction Internet is a virtual society.
Similar to a real society internet has a number of communities.
Communities are social knowledge reservoirs.
An assessment system is required to query the knowledge base.
P2P systems are the best examples of a community on internet.
Social Networking Websites (SNW) are latest P2P application.
Applied the assessment system to SNWs.
Orkut & LinkedIn
Network Members User Base
Orkut 22,000,000Designed specifically for friends and family.
LinkedIn 6,000,000Designed for professionals and adults.
MySpace 54,000,000Used mostly for fun and blogging.
Sporzoo 2,000,000Real Estate Investors and Professionals.
SelectedMinds 1,000,000Corporate social networking.
Reputation Assessment System Overview
The reputation assessment system consists of two components RQS : Derives reputation between directly connected nodes.
Generic reputation quantification function 4 canonical classes of the function with real life correspondence
• Fading Memory Averaging • Memory Less Summation • Fading Memory without Opinion Credibility• Memory Less Averaging
RRS : Derives reputation between dispersed nodes
Reputation based social computing
Related Work Gupta, Judge, and M. Ammar proposed system is specifically for peer-to-
peer systems. The system calculates the reputation based on variables which are specific to P2P systems and hence the system cannot be applied in other interaction based environments.
Damiani, Vimercati et.al. even though they incorporate community into their reputation function they do not take into consideration the credibility of the opinion.
Ecommerce websites such as eBay©, Yahoo© auction and Amazon© use a feedback based reputation system which is a simple memory less summation system .
Marti and Garcia-Molina propose a reputation system for a peer-to-peer environment, but their threat model only considers one class of attacks in which a malicious group is trying to work against the system. Their system is a simple voting system.
Reputation Assessment System Structure
Reputation Quantification System
Reputation Reasoning System
Notation
Reputation Quantification System
Estimates reputation
Any transaction consists of 3 sets Producer Product Consumer
With respect to reputation Producer: Opinion provider Product : Reputation Consumer : Individuals who use the
reputation to make decisions.
Opinion (O) & Time of Opinion (T)
R R
R
R
R
R
R
R
R RProduct
ConsumerProducer
Group Rating (W) & Number (N)
Opinion (O) & Time of Opinion (T)
R R
R
R
R
R
R
R
R RProduct
ConsumerProducer
Group Rating (W) & Number (N)
Opinion (O) & Time of Opinion (T)
R R
R
R
R
R
R
R
R RProduct
ConsumerProducer
Group Rating (W) & Number (N)
Opinion (O) & Time of Opinion (T)
R R
R
R
R
R
R
R
R RProduct
ConsumerProducer
Group Rating (W) & Number (N)
Social Factors Impacting Individuals Reputation
What does someone has to say about the individual – opinion
What is social standing of that someone – reputation
How long ago was this opinion expressed – age of opinion
How many people are saying something about the individual – number
What kind of group of people is expressing these opinions – group rating
All these factor not considered all the time some are omitted – impact
variable
Some may have more effect than others – impact weight
Generic Reputation Function (GRF) imp
Factors Influencing Reputation Computation Opinion (O) Reputation of Opinion Provider (R) Age of the Opinion (T) Number of Transactions (N) Group Reputation (W) Impact Parameters
Impact Variable (X) Impact Weight (α)
Stabilizing Factors Decay Rate (λ) Stable Value (Φ)
eeWN
eORWR jT
jWWNN
TXTj
OORR
m
j
N
j
TXX
N
j
TX
j
X
jm
k kAt
1 1
)(
1
)(
1
Different Scenarios
Are all variable (R,T,N,W) are valid in all scenarios ?
Table shows the 4 main scenario examples which have real life correspondence.
1110University ~ Students
0101University ~ Faculty
1110Automobile ~ Buyer
1111Automobile ~ Mechanic
1001Automobile ~ Manufacturer
1110Satellite ~ Satellite Service User
0101Satellite ~ Space Agency
1110Protocol ~ Users
1111Protocol ~ Companies
1111Course Material ~ Other Teachers
1001Course Material ~ Preparing Teachers
1110Course Material ~ Student
1111Article ~ Reader
0101Article ~ Journal
1001Article ~ Writer
0111Article ~ Reviewer
0111Movie ~ Critics
1110Movie ~ Viewers
1001Book ~ Author
1111Book ~ Reader
WNTRTarget ~ Evaluator
Canonical Classes of GRF GRF can be applied to different environments.
Every environment has different requirements.
Table shows the various real life combinations possible for the customizable variables
There are four broad classes of the GPF Fading Memory Averaging Memory Less Summation Fading Memory without Opinion Credibility Memory Less Averaging
Target ~ Evaluator R T N
Book ~ Reader 1 1 1
Book ~ Author 1 0 0
Movie ~ Viewers 0 1 1
Article ~ Journal 1 0 1
A Fading Memory Averaging Function
Valid Variables
Reader’s individual Reputation (R)
Reader’s opinion (O)
How fresh is the Opinion (T)
Number of Readers
nT
N
j
Tj
N
j
TjX
j
X
j
Ae
eeOR
RXTTOORR
t
1
)(
1
)(
)(
Target ~ Evaluator R T N
Book ~ Reader 1 1 1
A Memory Less Averaging Function
Valid Variables
Faculty’s Opinion (O)
Faculty’s Reputation (R)
Number of Opinions (N)
NeOR
RN
j
TjX
j
X
j
A
TOORR
t 1
)( 0
)(
Target ~ Evaluator R T N
University ~ Faculty 1 0 1
A Fading Memory Averaging Function without Opinion Credibility
Valid Variables
Viewer’s Opinion (O)
Age of Opinion (T)
Number of Opinions (N)
nT
N
j
Tj
N
j
TjX
jj
Ae
eeOR
RXTTOOR
t
1
)(
1
)(0
)(
Target ~ Evaluator R T N
Movie ~ Viewers 0 1 1
A Memory Less Summation Function
Valid Variables
Authors Opinions (O)
Authors Reputation (R)
NeOR
RN
j
TjX
j
X
j
A
TOORR
t0
0
)( 1
)(
Target ~ Evaluator R T N
Book ~ Author 1 0 0
Notation
Ei – Unique entity in super set E
Entities form various sets (A,B,C…)
Ei belongs to various sets.
AB denotes existence of relationship
between A & B.
Mr = A X B denotes relationship matrix
between the set members.
Mri – represents relationships of set A’s
members
MIj – represents relationships of B’s jth
member with all members of A
E1
E2
E4
E3
E
A BE4
Relationship Reasoning System Set of operations on relationship matrix M Relationship operations
Equivalence : M = N
Reflection : R = MT
Synthesis : S = M x N
Intersection : E = M ∩ N
Union : U = M U N
Exclusion : X = M θ N
Dediagonization :
Quantization :
Set Operations Column Extraction : (ψ)
Row Extraction : (ρ)
Max Row : (ξ)
Max Column : (Φ)
Zero Column :(θ)
M̂
M
Threats To Reputation Function
Attackers : Always lie Average Group
Good Group
Bad Group
Evaluators : Always honest
Target
Offender : Mostly honest but sometimes lies.
Reputation Attacks Classification
Attacks have been classified into 4 types
One – One Vendetta - Fight between two individuals
Many – One Gang Attack - A group working against an
individual
One – Many Dr Jekyll & Mr. Hyde - Dual behavior
Many – Many Mutual Boosting - Two-attacker groups join
together to mutually inflate reputations.
One Many
One Vendetta Dr Jekyll & Mr. Hyde
Many Gang Attack Mutual Boosting
Experimental Evaluation
Vendetta
Single target and single attacker
Attacker has random or high personal Reputation
Attacking frequency is random
Graphs - final Reputation against time of opinion
Personal attack has a very limited or no damaging effect
Fading Memory Averaging Function graphs
AT
TG
EV
EV
EV
Vendetta Graphs 1
0.97
0.15
0.97
0.520.99
0
0.2
0.4
0.6
0.8
1
1.2
1 6 11 16 21 26 31 36 41 46 51 56 61 66 71 76 81 86 91 96
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
Evaluator Reputation
Attacker Opinion
Attacker Reputation
Final Reputation
Evaluator Opinion
Evaluator Reputation High
Evaluator Opinion Random
Attacker Reputation High
Attacker Opinion Low
Behavior of Reputation function when Attacker and Evaluators have high personal reputation
Attack Period
Vendetta Graphs 2
Evaluator Reputation Random
Evaluator Opinion High
Attacker Reputation Random
Attacker Opinion Low
Behavior of Reputation function when Attacker has low reputation and evaluators have random reputation
0.96
0.76
0.39
0.40
0.19
0
0.2
0.4
0.6
0.8
1
1 6 11 16 21 26 31 36 41 46 51 56 61 66 71 76 81 86 91 96
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
Evaluator Opinion
EvaluatorReputation
Final Reputation
Attacker Opinion
AttackerReputation
Attack Period
Gang Attack
Group attacking a single target
Attack group has random or high personal Reputations.
Attackers manage to bring down the reputation.
But no permanent damage.
Function recovers itself.
Fading Memory Averaging Function graphs
AT
AT
AT
TGEV
EV
EV
Gang Attack Graph 1
0
0.2
0.4
0.6
0.8
1
1 4 7 10 13 16 19 22 25 28 31 34 37 40 43 46 49 52 55 58 61 64 67 70 73 76 79 82 85 88 91 94 97 100 103 106 109
0.05
0.25
0.45
0.65
0.85
1.05
1.25
1.45
1.65
1.85 Evaluator Reputation
Attacker Opinion
Attacker Reputation
Final Reputation
Evaluator Opinion
Evaluator Reputation High
Evaluator Opinion High
Attacker Reputation High
Attacker Opinion Low
Behavior of reputation function when attackers and evaluators have high reputation
Attack Period
Gang Attack Graph 2
0.91
0.74
0.12
0.69
0.97
0
0.2
0.4
0.6
0.8
1
1 5 9 13 17 21 25 29 33 37 41 45 49 53 57 61 65 69 73 77 81 85 89 93 97 101 105 109
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2 Evaluator Opinion
AttackerReputation
Attacker Opinion
Final Reputation
EvaluatorReputation
Evaluator Reputation High
Evaluator Opinion High
Attacker Reputation Random
Attacker Opinion Low
Behavior of reputation function when attacker has random reputation and evaluators have high reputation
Attack Period
Dr. Jekyll & Mr. Hyde Dual conflicting behavior of a peer. The peer develops a high reputation, then misbehaves Evaluator group penalizes him. Peer’s reputation goes down. Recovers back to original value if peer indulges in honest
transactions.
AT
TG
TG
TG
Dr Jekyll & Mr. Hyde Graph 1
0.11
0.74
0.9
0
0.2
0.4
0.6
0.8
1
1.2
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49 51 53 55 57 59 61 63 65 67 69 71 73 75 77
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
EvaluatorOpinion
FinalReputation
EvaluatorReputation
Evaluator Reputation High
Normal period Evaluator Opinion High
Penalty period Evaluator Opinion Low
Behavior of reputation function when evaluators have high reputation
Penalty Period
Dr Jekyll & Mr. Hyde Graph 2Behavior of reputation function when evaluators have random reputation
Evaluator Reputation Random
Normal period Evaluator Opinion High
Penalty period Evaluator Opinion Low
0
0.2
0.4
0.6
0.8
1
1.2
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49 51 53 55 57 59 61 63 65 67 69 71 73 75 77
EvaluatorOpinion
EvaluatorReputation
FinalReputation
Penalty Period
Mutual Boosting
Groups intentionally give high opinion to each other
High reputed group’s reputation drops
Low reputed group’s reputation rises
Eventually all the groups have same reputation.G1
G1 G1
G1 G1
G2
G2 G2
G2 G2
Mutual Boosting GraphBehavior of reputation function when one group has high reputation, one has average reputation and one has low reputation members
Reputation Opinion
Member Group 1 High High
Member Group 2 Average High
Member Group3 Low High
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
1 2 3 4 5 6 7 8 9 10
MEMBER 0
MEMBER 1
MEMBER 2
Social Network Based Computations
Social Profile Mining
Social Fabric Analysis
Social Linkage Analysis
Social Ranking Analysis
Placement Within a Community
A B C D
Social Profile Mining
A B
C
D
Social Fabric Analysis
A B
C
D
Social Linkage Analysis
A B
C
D
Social Ranking Analysis
A B C D
Social Profile Mining
A B C D
Social Profile Mining
A B
C
D
Social Fabric Analysis
A B
C
D
Social Fabric Analysis
A B
C
D
Social Linkage Analysis
A B
C
D
Social Linkage Analysis
A B
C
D
Social Ranking Analysis
A B
C
D
Social Ranking Analysis
Social Profile Mining Definition : The process of searching large volumes of data for hidden patterns, to
predict future behavior
Uses: Geo-Profiling, Demographic information, Market Analysis.
SNW are a rich database of individual’s personal and professional information.
This data can be subjected to descriptive and inferential statistical analysis.
Descriptive: Mean, Mode & Standard Deviation
Inferential: Hypothesis testing, Estimation, Correlation, Regression.
Examples: Mining the cuisine and favorite movie sections of Orkut
Social Fabric Analysis
Definition: Given a weighted graph G(V,E), and given further one element v of V, find the strength of social property S on v for each v’ of V.
Social Fabric analysis derives social properties using primary relationships.
Primary relationships: father, mother, friend etc
Social properties: Influence, status, reputation, trust.
Example: Influence Propagation
Influence
Definition:” the act or power of producing an effect without apparent exertion of force or direct exercise of command”
Factors affecting influence: I-factors
Age Difference (F1)
Age determined from profile on SNW
Proximity of Nodes (F2)
City , State and Country determined from SNW profile
Relationship Type (F3)
Monitoring Interactions
Contact Frequency (F4)
Interaction Monitoring
jicountryjistatejicityfjiF .,,,,,
2
i,jjiF Relation,3
N
jiji
NkF
,scrap, 1
4
ageageF jiji
,1
Influence Context (F5)
Influencing node’s Reputation (F6)
Using RQS
Trust (F7)
Using RQS
Influence (I)
Tn
N
j
Tj
N
j
TjX RR
j
X
j ee
eORF
XTTOO
A
1
)(
1
)(
6
Tn
N
m
mT
N
m
TjX
mX
ie
eeKRF
XTTOORR
ji
1
)(
1
)(
7,
7654321 ,,,,,, FFFFFFFfI
Influence: Algorithmic Structure
Depth first traversal of network
Go to next node if relationship link
part of favorable links
Next node should satisfy I-factors to
continue traversal
Traversal length dependant on
relationship chain
Mostly 3 hops
Time complexity is O(|v| + |e|)
Influence (G, u, δ, λ) u - is the source node
Stack S = { };
Stack O = { };
Boolean x, y;
pathString (u) = null;
Push S, u;
while ( S is not empty ) do
u := Pop S;
Push S, u;
for each vertex v adjacent to u
if ( AgeV > 18 ) && ( F1(u, v) > 2) && ( F6(v) > δ) && (F7(u, v) > λ )
I = (AgeV /100) * (F1 (u, v)/50) * F6 (v) * F7 (u, v);
pathString(v) = Concatenate ( pathString(u), F3(u,v));
Push S, v;
StringMatching(pathString(v), patternDb())
I = I * patternDb(pathString(v)).value();
return I;
end if
end while
Influence Derivation
Peter
George
Bob
Laura
Kallis
JoeMartin
Friend
Spouse
Father
Coworker
Friend
Neighbor
Scraps
City, State, Country
Age
Scraps
City, State, Country
Age
Scraps
City, State, Country
Age
Scraps
City, State, Country
Age
Scraps
City, State, Country
Age
Scraps
City, State, Country
Age
Scraps
City, State, Country
Age
Peter
George
Bob
Laura
Kallis
JoeMartin
Friend
Spouse
Father
Coworker
Friend
Neighbor
Scraps
City, State, Country
Age
Scraps
City, State, Country
Age
Scraps
City, State, Country
Age
Scraps
City, State, Country
Age
Scraps
City, State, Country
Age
Scraps
City, State, Country
Age
Scraps
City, State, Country
Age
George’s Social Network On Orkut
George
D
M
K
J
H
G
B
I
A
Y
L
N
ZN
Peter
George
Bob
Laura
Kallis
JoeMartin
Spouse
Father
Coworker
Friend
Neighbor
Scraps
City, State, Country
Age
Scraps
City, State, Country
Age
Scraps
City, State, Country
Age
Scraps
City, State, Country
Age
Scraps
City, State, Country
Age
Scraps
City, State, Country
Age
Scraps
City, State, Country
Age
0.9
0.9
0.5
0.2
0.50.20.4
0.7
0.02
0.1
Peter
George
Bob
Laura
Kallis
JoeMartin
Spouse
Father
Coworker
Friend
Neighbor
Scraps
City, State, Country
Age
Scraps
City, State, Country
Age
Scraps
City, State, Country
Age
Scraps
City, State, Country
Age
Scraps
City, State, Country
Age
Scraps
City, State, Country
Age
Scraps
City, State, Country
Age
0.9
0.9
0.5
0.2
0.50.20.4
0.7
0.02
0.1
Numbers indicate F7 values
Social Linkage Analysis
Definition: Given a graph G(V,E) , where each edge u,v denotes a relationship r(u,v), we want to find the most effective relationship chain from source s to the sink t subject to certain constraints
Source “s” and Sink “t” are fixed
Determine the relationship chain that maximizes probability of contact between s and t
Example : Vested Socialite
Vested Socialite
Definition:” A person trying to force himself into other people’s social
network in order to achieve some objective”
Factors for social insertion: V-factors
Objective
Constraining factors
Direct influence
Derived influence
Relationship chain
Hate list
Vested Socialite : Algorithmic Structure
Determine path from s to t
Drop path if it contains hate list member
Drop path if it violates constraints.
If more than 1 path , chose the one with
highest contact probability
Some favorable relationship chains:
Friend{1,}
Friend-Relative-Friend
Relative-Friend{1,2}
Relative-Friend-Relative
Coworker-Friend,
Coworker-Relative
FindPath(G, start, t, path)
Array paths;
Concatenate (path, start)
If start == end
return path
for each node n in G
if(StringCompare(F3(start, n), Enemy)) == false
if n not in path
newpath=FindPath(G, n, t, path)
paths.add(newpath)
end if
end if
return paths[ ];
InsertProbabilty(paths[ ], start, t)
for each path p in paths
for each node n in path p
rchain(p) = rchain(p) + F3(n, n+1)
StringMatch(rchain(p), patternDb())
if true then
insertprobability = patternDb(relationchain(p)).value();
return insertprobabilty;
break;
Vested Socialite
John
Victor
Gary
DavidSteven
Paul Raj
Frank
John
Victor
Gary
DavidSteven
Paul Raj
Frank
John is “s” and Victor is “t”
John
Victor
Gary Steven David Frank
Paul Raj
ENEMY
SUBORDINATE
FRIEND
FATHER
FATHER
BOSS
FATHER
FRIEND
FRIEND
FRIENDFRIEND
John
Victor
Gary
DavidSteven
Paul Raj
Frank
Father
Friend
Enemy
Friend
Father
Subordinate
FatherBoss
Friend
Friend
Friend
John
Victor
Gary
DavidSteven
Paul Raj
Frank
Father
Friend
Enemy
Friend
Father
Subordinate
FatherBoss
Friend
Friend
Friend
Favorable paths in blue. Unfavorable in red
Social Ranking Analysis
Definition: Given a graph G(V,E), we want to determine the ranking of all v
of V based on a social property P.
Ranking algorithms are used while making decisions
Ranking changes with social property.
Example: Ranking based on Trust
Trust Based Ranking
Two groups are considered
Sink ranking group Nodes that trust the target “t”
Source ranking group Nodes the target “t” trusts
Final ranking is combination of source and sink rankings.
j1
i
j4
j3j2
Tj1i
Tj2iTj3i
Tj4i
j1
i
j4
j3j2
Tj1i
Tj2iTj3i
Tj4i
j1
i
j4
j3j2
Tij1
Tij2Tij3
Tij4
j1
i
j4
j3j2
Tij1
Tij2Tij3
Tij4
Algorithm: Eigen Computation
Sink ranking
Source ranking
Generalized Equation
λ is an eigenvalue of T.
The eigenvalue determines the eigenvectors.
The eigenvectors represent the rankings.
Eigenvalue and eigenvector computation require O(n3) operations
xTx ji
n
jji
1
1
xTy jiji
n
j
T
1
1
xxTTT
Community Placement
Definition: Community placement is the problem of finding highly connected individuals within a community who could influence a majority of the community to agree on or believe in an issue.
Propagandist use a brute force method – meeting every individual
Better approach is to influence a coterie
Extension of social ranking with the extra task of finding the appropriate individuals.
Multi-Faith Group
Gather a group of individuals from varied faiths and use them to propagate religious tolerance.
The individuals should satisfy the following conditions Should be from different faiths Should be highly connected Should have friends belonging to different faiths.
Multi-Faith Group: Algorithm
Two main stages: ranking and propagation
Ranking algorithm finds everyone’s ranking.
Top ranked individual sends agenda to his neighbors and the influencing power is recorded.
He is then pruned from the graph.
The above step is repeated until a majority is influenced
Time complexity would be O(n3) + O(|V|+|E|)
Conclusion
We have presented a framework to represent and reason with
general case of social relationship network in a formal way.
We have provided an assessment system using which the
strength of relationship can be measured.
We provide a classification for the potential applications of the
rich database that users of social networking websites are
unknowingly creating and further enriching everyday.
The application classifications are substantiated by providing
complex examples for each class.
References
Gupta, M., Judge, P., and Ammar, M.,( 2003) "A reputation system for peer-to-peer networks" in NOSSDAV. pp 144 -152
Damiani, E., De Capitani di Vimercati, S., Paraboschi, S.,and Samarati, P., (2003)“Managing and sharing servants’’ reputations in P2P systems”. IEEE Transactions on Data and Knowledge Engineering.pp 840–854.
Marti, S., and Garcia-Molina, H.,(2004). “Limited reputation sharing in P2P systems”. In Proc. of the 5th ACM conference on Electronic commerce, New York, NY, USA, pp 91-101,
Gambetta, D., (1988) “Trust making and breaking cooperative relations”. New York: Blackwell
McKnight, D., & Chervany, N., (1996) “The Meaning of Trust”. University of Minnesota MIS Research Center Working Paper Series, WP 96-104
McShane, S., (1995). Canadian Organization Behavior. Ottawa:Irwin. Meneses, J., (2004) “The Orkut.com case: a reflection on the exploration of new ways
to online sociability in the tradition of the study of virtual communities”. Xiong, L., and Liu, L.,. (2003). “A reputation-based trust model for peer-to-peer
ecommerce communities”. IEEE Conference on E-Commerce (CEC'03). pp 275 – 281
Malaga., R. A., (2001.). “Web-based reputation management systems: Problems and suggested solutions”. Electronic Commerce Research. pp 403 – 417
THANK YOU !!