Download as pdf or txt
Download as pdf or txt
You are on page 1of 25

Positional analysis of multi-mode

fuzzy networks

Miroslav Ćirić
Department of Computer Science
Faculty of Sciences
University of Niš, Serbia
miroslav.ciric@pmf.edu.rs

joint work with

Jelena Ignjatović, Ivan Stanković

5th Conference on Information Theory and Complex Systems – TINKOS 2017


November 9th – 10th, 2017, Belgrade, Serbia

1 Miroslav Ćirić, Jelena Ignjatović, Ivan Stanković Positional analysis of multi-mode fuzzy networks
Social network analysis. Positional analysis

Social network analysis


⋆ a branch of sociology and mathematics which provides formal models
and methods for the systematic study of social structures
⋆ social network: set of actors and ties between them
⋆ social networks share many common properties with other networks
⋆ methods of social network analysis are applied to the analysis of networks in
general – many types of networks in computer science, physics, biology, etc.:
the hyperlink structure on the Web, the electric grid, computer networks,
information networks, various large-scale networks appearing in nature, etc.
⋆ fuzzy approach – social relations are essentially fuzzy

Positional analysis
⋆ identify the position or role of actors in the network on the basis of
relationships between them
⋆ example: teroristic group – identify roles (leaders, etc.) on the basis on
communication between the group members

2 Miroslav Ćirić, Jelena Ignjatović, Ivan Stanković Positional analysis of multi-mode fuzzy networks
Structure of truth/membership values

Complete residuated lattice


ordered algebraic structure L = (L, ∧, ∨, ⊗, →, 0, 1) such that
(L1) (L, ∧, ∨, 0, 1) is a complete lattice with the least element 0 and the
greatest element 1,
(L2) (L, ⊗, 1) is a commutative monoid with the unit 1,
(L3) ⊗ and → satisfy the residuation property:
x ⊗ y 6 z ⇔ x 6 y → z.

Interpretations
⋆ ⊗ – strong conjunction; ∧ – weak conjunction; ∨ – (weak) disjunction;
⋆ → – implication (residual implication, residuum)
V
⋆ infimum (infinite) – universal quantifier
W
⋆ supremum (infinite) – existential quantifier
⋆ ordering relation 6 – semantic consequence relation
⋆ residuation property – Deduction Theorem
3 Miroslav Ćirić, Jelena Ignjatović, Ivan Stanković Positional analysis of multi-mode fuzzy networks
Fuzzy sets and fuzzy relations

Fuzzy sets
⋆ fuzzy subset of a set A is a function α : A → L
⋆ equality: α = β ⇔ α(a) = β(a), for each a ∈ A
⋆ inclusion: α 6 β ⇔ α(a) 6 β(a), for each a ∈ A
⋆ union and intersection
_  _ ^  ^
αi (a) = αi (a), αi (a) = αi (a)
i∈I i∈I i∈I i∈I

⋆ if A is finite, with |A| = n, then α is an n-dimensional fuzzy vector


⋆ LA – the set of all fuzzy subsets of A

Fuzzy relations
⋆ fuzzy relation between sets A and B is a fuzzy subset of A × B, i.e., a
function R : A × B → L
⋆ if A and B are finite, |A| = m and |B| = n, then R is a m × n fuzzy matrix
⋆ LA×B – the set of all fuzzy relations between A and B

4 Miroslav Ćirić, Jelena Ignjatović, Ivan Stanković Positional analysis of multi-mode fuzzy networks
Fuzzy equivalences and fuzzy quasi-orders
a fuzzy relation R ∈ LA×A is
⋆ reflexive, if R(a, a) = 1, for all a ∈ A;
⋆ symmetric, if R(a, b) = R(b, a), for all a, b ∈ A;
⋆ transitive, if R(a, b) ⊗ R(b, c) 6 R(a, c), for all a, b, c ∈ A.

Fuzzy equivalence
⋆ reflexive, symmetric and transitive fuzzy relation
⋆ for a fuzzy equivalence E on A and a ∈ A, a fuzzy subset Ea ∈ LA defined
by Ea (b) = E(a, b) is an equivalence class of E determined by a

Fuzzy quasi-order
⋆ reflexive and transitive fuzzy relation
⋆ for a fuzzy quasi-order Q on A and a ∈ A, a fuzzy subset aQ ∈ LA defined
by aQ(b) = Q(a, b) is an afterset of Q determined by a
⋆ aftersets – columns, principal upper sets (principal filters)
⋆ dual concept: foreset Qa ∈ LA defined by Qa(b) = Q(b, a)

5 Miroslav Ćirić, Jelena Ignjatović, Ivan Stanković Positional analysis of multi-mode fuzzy networks
Compositions

Composition of fuzzy relations (matrix product)


⋆ for R ∈ LA×B and S ∈ LB×C , the composition R ◦ S ∈ LA×C is defined by
_
(R ◦ S)(a, c) = R(a, b) ⊗ S(b, c)
b∈B

Composition of a fuzzy set and a fuzzy relation (vector-matrix


products)
⋆ for R ∈ LA×B , α ∈ LA and β ∈ LB , the compositions α ◦ R ∈ LB and
R ◦ β ∈ LA are defined by
_ _
(α ◦ R)(b) = α(a) ⊗ R(a, b), (R ◦ β)(a) = R(a, b) ⊗ β(b)
a∈A b∈B

Composition of fuzzy sets (scalar product, dot product)


⋆ for α, β ∈ LA , the composition α ◦ β ∈ L is defined by
_
α◦β = α(a) ⊗ β(a)
a∈A

6 Miroslav Ćirić, Jelena Ignjatović, Ivan Stanković Positional analysis of multi-mode fuzzy networks
Residuals

⋆ right residual of T by S (for S ∈ LA×B , T ∈ LA×C ): S\T ∈ LB×C is given by


^
(S\T)(b, c) = S(a, b) → T(a, c)
a∈A

⋆ left residual of T by S (for S ∈ LB×C , T ∈ LA×C ): T/S ∈ LA×B is given by


^
(T/S)(a, b) = S(b, c) → T(a, c)
c∈C

⋆ residuation property: S ◦ T 6 U ⇔ T 6 S\U ⇔ S 6 U/T


⋆ very important in solving inequalities and equations with fuzzy relations
⋆ S\U = max {X ∈ LB×C | S ◦ X 6 U}, U/T = max {X ∈ LA×B | X ◦ S 6 U}
⋆ other kinds of residuals:
fuzzy set by scalar fuzzy set by fuzzy set
scalar by fuzzy set fuzzy set by fuzzy relation

7 Miroslav Ćirić, Jelena Ignjatović, Ivan Stanković Positional analysis of multi-mode fuzzy networks
Fuzzy social networks

One-mode fuzzy network – O = (A, R)


⋆ A – non-empty set, R = {Ri }i∈I ⊂ LA×A

Two-mode fuzzy network – T = (A, B, R)


⋆ A, B – two different non-empty sets, R = {Ri }i∈I ⊂ LA×B

Multi-mode fuzzy network – M = (A1 , . . . , An , R)


⋆ A1 , . . . , An – non-empty sets
⋆ R – system of fuzzy relations between Aj and Ak defined for some pairs (j, k)
⋆ formally:
• J ⊆ [1, n] × [1, n] such that (∀j ∈ [1, n])(∃k ∈ [1, n]) (j, k) ∈ J or (k, j) ∈ J
• {Ij,k }(j,k)∈J – collection of non-empty sets
j,k j,k
• R = {Ri | (j, k) ∈ J, i ∈ Ij,k }, Ri ∈ LAj ×Ak , for all (j, k) ∈ J and i ∈ Ij,k
⋆ complex synthesis of one-mode and two-mode fuzzy networks

8 Miroslav Ćirić, Jelena Ignjatović, Ivan Stanković Positional analysis of multi-mode fuzzy networks
Regular fuzzy equivalences. Blockmodeling

Regular fuzzy equivalences on O = (A, R)


⋆ fuzzy equivalences that are solutions to the system Ri ◦ X = X ◦ Ri , (i ∈ I)
⋆ play the key role in the positional analysis – White, Reitz (1983)
⋆ informally: two actors are considered to be regularly equivalent if they are
equally related to equivalent others

Blockmodeling
⋆ blockmodeling – related partitioning of the network
⋆ data reduction method which clusters actors who have substantially similar
patterns of relationships with others and interprets the pattern of relationships
among the clusters
⋆ blockmodel image – structural summary of the original complex network

Ignjatović, Ćirić, Bogdanović (Fuzzy Sets and Systems, 2010)


Ignjatović, Ćirić (Filomat, 2012)

9 Miroslav Ćirić, Jelena Ignjatović, Ivan Stanković Positional analysis of multi-mode fuzzy networks
Two-mode networks

Two-mode networks
⋆ affiliation or bipartite networks
⋆ examples: people attending events, organizations employing people, justices
on a court rendering decisions, nations belonging to alliances, etc.

Positional analysis of two-mode networks


⋆ identify positions in both modes of the network
⋆ indirect approach: reduction to one or two one-mode networks –
computationally demanding, loss of information
⋆ direct approach: solving two-mode systems of fuzzy relation equations and
inequalities

Stanković, Ćirić, Ignjatović (Fuzzy Sets and Systems, 2017)

10 Miroslav Ćirić, Jelena Ignjatović, Ivan Stanković Positional analysis of multi-mode fuzzy networks
Multi-mode networks – examples

Simple organization network

11 Miroslav Ćirić, Jelena Ignjatović, Ivan Stanković Positional analysis of multi-mode fuzzy networks
Multi-mode networks – examples
Citations

Papers

A
p
hi

pp
rs

ea
ho

ra
ut

nc
A

e
Publishing

Semantics
Researchers Words
Ed
iti
ng

Journals,
proceedings

Network of academic publications


12 Miroslav Ćirić, Jelena Ignjatović, Ivan Stanković Positional analysis of multi-mode fuzzy networks
Multi-mode networks – examples

Network of the elite of cancer researchers in France


13 Miroslav Ćirić, Jelena Ignjatović, Ivan Stanković Positional analysis of multi-mode fuzzy networks
Multi-mode networks – examples

Another view on the same network


14 Miroslav Ćirić, Jelena Ignjatović, Ivan Stanković Positional analysis of multi-mode fuzzy networks
Multi-mode networks – examples

Genetic regulatory (interaction) network

15 Miroslav Ćirić, Jelena Ignjatović, Ivan Stanković Positional analysis of multi-mode fuzzy networks
Multi-mode systems of FRIE
M = (A1 , . . . , An , R) – multi-mode fuzzy network

j,k j,k
αj ◦ Ri = Ri ◦ αk , (j, k) ∈ J, i ∈ Ij,k , (1)
j,k j,k
αj ◦ Ri 6 Ri ◦ αk , (j, k) ∈ J, i ∈ Ij,k , (2)
j,k j,k
αj ◦ Ri > Ri ◦ αk , (j, k) ∈ J, i ∈ Ij,k , (3)

⋆ α1 , . . . , αn – unknowns, αj takes values in LAj ×Aj , for each j ∈ [1, n]


⋆ solutions – n-tuples of fuzzy relations from LA1 ×A1 × · · · × LAn ×An
⋆ n-tuples are ordered coordinatewise, we read η 6 θ as ”η contained in θ”
⋆ (1), (2) and (3) – two-sided linear equations/inequalities

⋆ solutions of (1) – regular n-tuples w.r.t. M


⋆ solutions of (2) and (3) – left regular and right regular n-tuples w.r.t. M

16 Miroslav Ćirić, Jelena Ignjatović, Ivan Stanković Positional analysis of multi-mode fuzzy networks
Existence of the greatest solutions

Theorem 1 (Existence of the greatest solutions)

Let (̺10 , ̺20 , . . . , ̺n0 ) ∈ LA1 ×A1 × · · · × LA2 ×A2 × LAn ×An .

⋆ Sets of all solutions of systems (1), (2) and (3) form complete lattices.
⋆ There exist the greatest solutions contained in (̺10 , ̺20 , . . . , ̺n0 ).

⋆ If ̺10 , ̺20 , . . . , ̺n0 are fuzzy quasi-orders, all components of the greatest solutions
are also fuzzy quasi-orders.

To ensure that all components are fuzzy equivalences

⋆ new equations/inequalities are added with α−1


j
instead of αj

⋆ (̺10 , ̺20 , . . . , ̺n0 ) is taken to be an n-tuple of fuzzy equivalences

17 Miroslav Ćirić, Jelena Ignjatović, Ivan Stanković Positional analysis of multi-mode fuzzy networks
Computation of the greatest solutions
Theorem 2 (Computation of the greatest solution)

Let (̺10 , ̺20 , . . . , ̺n0 ) ∈ LA1 ×A1 × LA2 ×A2 × · · · × LAn ×An .
Let {(̺1r , ̺2r , . . . , ̺nr )}r∈N be a decreasing sequence in LA1 ×A1 × LA2 ×A2 × · · · × LAn ×An
defined by

(̺11 , ̺21 , . . . , ̺n1 ) = (̺10 , ̺20 , . . . , ̺n0 ),


^ ^  ^ ^ k,j
j,k j,k k,j

̺jr+1 = ̺jr ∧ (Ri ◦ ̺kr )/Ri ∧ Ri \(̺kr ◦ Ri ) , j ∈ [1, n].
k∈Λj i∈Ij,k k∈Pj i∈Ik,j

(a) if there exists s ∈ N such that (̺1s , ̺2s , . . . , ̺ns ) = (̺1s+1 , ̺2s+1 , . . . , ̺ns+1 ), then
(̺1s , ̺2s , . . . , ̺ns ) is the greatest solution of (1) contained in (̺10 , ̺20 , . . . , ̺n0 );

(b) if A1 , A2 , . . . , An are finite sets and the subalgebra L(R ∪ {̺10 , ̺20 , . . . , ̺n0 })
satisfies DCC, then the sequence {(̺1r , ̺2r , . . . , ̺nr )}r∈N is finite and there exists
s ∈ N such that (̺1s , ̺2s , . . . , ̺ns ) = (̺1s+1 , ̺2s+1 , . . . , ̺ns+1).

18 Miroslav Ćirić, Jelena Ignjatović, Ivan Stanković Positional analysis of multi-mode fuzzy networks
Computation of the greatest solutions (cont.)

Λ = {j ∈ [1, n] | (∃k ∈ [1, n]) (j, k) ∈ J} – αj appears on the left side


P = {j ∈ [1, n] | (∃k ∈ [1, n]) (k, j) ∈ J} – αj appears on the right side

Theorem 3 (Computation without iteration)

Suppose that Λ ∩ P = ∅, and let (̺10 , ̺20 , . . . , ̺n0 ) ∈ LA1 ×A1 × LA2 ×A2 × · · · × LAn ×An .
Let (̺1 , ̺2 , . . . , ̺n ) ∈ LA1 ×A1 × LA2 ×A2 × · · · × LAn ×An be defined by
^ ^
j,k j,k

̺j = ̺j0 ∧ (Ri ◦ ̺k0 )/Ri , j ∈ [1, n].
k∈Λj i∈Ij,k

Then (̺1 , ̺2 , . . . , ̺n ) is the greatest solution of (2) contained in (̺10 , ̺20 , . . . , ̺n0 ).

19 Miroslav Ćirić, Jelena Ignjatović, Ivan Stanković Positional analysis of multi-mode fuzzy networks
What if the sequence does not stabilize?

⋆ instances of multi-mode systems


⋆ consist of linear inequalities/equations
j,k j,k j,k j,k j,k j,k j,k j,k
αj ◦ Ri 6 Ri , Ri 6 αj ◦ Ri , Ri ◦ αk 6 Ri , Ri 6 Ri ◦ αk .

⋆ do not require iteration


⋆ generalize structural equivalences – Lorrain, White (1971)
⋆ in practice they have shown too restrictive

⋆ crisp solutions of multi-mode systems


⋆ we have provided efficient algorithms for their computing, which always work
⋆ however, fuzzy solutions give better results in practice

20 Miroslav Ćirić, Jelena Ignjatović, Ivan Stanković Positional analysis of multi-mode fuzzy networks
Blockmodeling – Quotient networks

Quotient networks
j,k
⋆ M = (A1 , A2 , . . . , An , R), R = {Ri | (j, k) ∈ J, i ∈ Ij,k } – multi-mode network
⋆ (̺1 , ̺2 , . . . , ̺n ) ∈ LA1 ×A1 × · · · × LAn ×An – n-tuple of fuzzy quasi-orders
⋆ ̺˜j = ̺j ∧ ̺j−1 – natural fuzzy equivalence of ̺j , A/̺˜j – the related quotient set
j,k j,k
⋆ R̃ = {R̃i | (j, k) ∈ J, i ∈ Ij,k }, R̃i ∈ L(A/̺˜j )×(A/̺˜k ) – well defined by
j,k j,k
R̃i ((̺˜j )aj , (̺˜k )ak ) = (̺j ◦ Ri ◦ ̺k )(aj , ak ), for all aj ∈ Aj and ak ∈ Ak .

⋆ M˜ = (A1 /̺˜1 , A2 /̺˜2 , . . . , An /̺˜n , R̃) – quotient network


The crisp case
⋆ (̺1 , ̺2 , . . . , ̺n ) is a solution of (2) if and only if
j,k j,k
((̺˜j )aj , (̺˜k )ak ) ∈ R̃i ⇔ (∀xj ∈ ̺j aj )(∃xk ∈ ̺k ak ) (xj , xk ) ∈ Ri .
⋆ (̺1 , ̺2 , . . . , ̺n ) is a solution of (3) if and only if
j,k j,k
((̺˜j )aj , (̺˜k )ak ) ∈ R̃i ⇔ (∀xk ∈ ak ̺k )(∃xj ∈ aj ̺j ) (xj , xk ) ∈ Ri .

21 Miroslav Ćirić, Jelena Ignjatović, Ivan Stanković Positional analysis of multi-mode fuzzy networks
Example 1
Grouping employees and jobs
⋆ two-mode network T = (A, B, R), R = {Ri }i∈I
⋆ A – the set of all employees of some company
⋆ B – the set of all jobs which this company performs for other companies
⋆ I set of these other companies
⋆ the jobs for the company i ∈ I are allotted to employees by a relation
Ri ⊆ A × B

⋆ Task: group employees into teams and jobs into groups of jobs so that
⊲ the teams and groups of jobs are as wide as possible
⊲ for any company, a group of jobs γ is assigned to a team θ if and only if
∗ for every employee from θ there is a job from γ which he has already
performed for that company
∗ for every job from γ there is an employee from θ who has already
preformed that job for that company

⋆ such grouping can be done using the regular pair of equivalences w.r.t. T

22 Miroslav Ćirić, Jelena Ignjatović, Ivan Stanković Positional analysis of multi-mode fuzzy networks
Example 2
The modified problem
⋆ the teams and the groups of jobs have wider and narrower parts
⋆ the narrower parts – the cores of the teams and groups of jobs
⋆ for any company, a group of jobs γ is assigned to a team θ if and only if
∗ for every employee from the core of θ there is a job from γ which he has
already performed for that company
∗ for every job from the core of γ there is an employee from θ who has already
preformed that job for that company
⋆ such grouping can be done using the regular pair (α, β) of quasi-orders
⋆ the wider teams and groups of jobs are the aftersets of α and the foresets of β
⋆ the narrower teams and groups of jobs are equivalence classes of the natural
equivalences α̃ and β̃
⋆ the core of the team performs the main part of the assigned jobs, and the rest of the team
assists the core in the jobs that they have not previously performed and in other cases
when they need help
⋆ the core of the group of jobs assigned to the team are main jobs they have to perform,
while the rest of this group are those jobs for which the members of the team could be
engaged to assist

23 Miroslav Ćirić, Jelena Ignjatović, Ivan Stanković Positional analysis of multi-mode fuzzy networks
Example 3
Adding a third mode
⋆ third mode: skills, e.g., knowledge of specific software packages, if the
considered company is a software company
⋆ groups of software packages assigned to teams
⋆ for any company, a group of software packages π is assigned to a team θ
if and only if
∗ for every employee from θ there is a software package from π for which the
employee is qualified
∗ for every software package from π there is an employee from θ who is
qualified for that software package

⋆ such grouping can be done using the regular triples of equivalences


⋆ version with cores – regular triples of quasi-orders

24 Miroslav Ćirić, Jelena Ignjatović, Ivan Stanković Positional analysis of multi-mode fuzzy networks
Thanks!

Thank you
for your attention!

25 Miroslav Ćirić, Jelena Ignjatović, Ivan Stanković Positional analysis of multi-mode fuzzy networks

You might also like