In this talk, I present recent results on the measurement of conflict in
signed social network. A signed social network is a network in which
both positive and negative ties are present, for instance representing
friendship and enmity, or trust and distrust. Such networks have long
been studied under the aspect of balance theory, i.e., considering
whether the individuals can be grouped into two groups, such that the
sign of all ties reflects the partition, or equivalently in terms of the
individual configuration of tryads. The failure of such a structure to
be present in a signed social network is usually designated as conflict,
and measuring the amount of conflict in a given signed social network is
an open problem. In this work, I present a novel measure of conflict,
based on algebraic graph theory, and considering a previously known
variant of the Laplacian matrix for signed graphs. The talk will show
both theoretical motivations for the measure, as well as an evaluation
of its utility at signed social network analysis using multiple
real-world signed social networks.
Unveiling the Role of Social Media Suspect Investigators in Preventing Online...
Algebraic Graph-theoretic Measures of Conflict
1. Institute for Web Science and Technologies · University of Koblenz-Landau, Germany
Algebraic Graph-theoretic
Measures of Conflict
Jérôme Kunegis
Based on work performed in collaboration with Christian Bauckhage, Andreas
Lommatzsch, Stephan Spiegel, Jürgen Lerner, Fariba Karimi and Christoph Carl Kling
Journée Graphes et Systèmes Sociaux (JGSS), March 18, 2016, Avignon
4. Algebraic Graph-theoretic Measures of Conflict 4Jérôme Kunegis
The Slashdot Zoo
Slashdot Zoo: Tag users as friends and foes
Graph has two types of edges: friendship and enmity
5. Algebraic Graph-theoretic Measures of Conflict 5Jérôme Kunegis
Signed Directed Social Network
me
FriendFoe
FanFreak
The resulting graph is sparse,
square, asymmetric and has
signed edge weights.
You are the fan of your friends and the freak of your foes.
13. Algebraic Graph-theoretic Measures of Conflict 13Jérôme Kunegis
Measuring Tryadic Conflict
Definition:
C =₃
#negTriangles
#totalTriangles
(cf. Signed relative clustering coefficient, Kunegis et al. 2009)
C = 0₃
14. Algebraic Graph-theoretic Measures of Conflict 14Jérôme Kunegis
Balance on Longer Cycles
Equivalent definitions: a graph is balanced when
(a) all cycles contain an even number of negative edges
(b) its nodes can be partitioned into two groups such
that all positive edges are within each group, and all
negative edges connect the two groups
15. Algebraic Graph-theoretic Measures of Conflict 15Jérôme Kunegis
Frustration
Definition: The minimum number of edges f that
have to be removed from a signed graph to make
the graph balanced.
Example: f = 1
16. Algebraic Graph-theoretic Measures of Conflict 16Jérôme Kunegis
Frustration (partitioning view)
Definition: minimum number of edges that are
frustrated (i.e., inconsistent with balance) given any
partition of the graph's nodes into two groups.
Example:
f = 1
17. Algebraic Graph-theoretic Measures of Conflict 17Jérôme Kunegis
Frustration: Computation
Computation of frustration is equivalent to
MAX-2-XORSAT
max (a XOR b) + (¬a XOR c) + (b XOR ¬d) ⋯
MAX-2-XORSAT is NP complete
Solution: Relax the problem
see overview in (Facchetti & al. 2011)
a
b
c
d
18. Algebraic Graph-theoretic Measures of Conflict 18Jérôme Kunegis
Algebraic Formulation
Let G = (V, E, σ) be a signed graph. σuv = ±1 is the
sign of edge (uv).
Given a partition V = S T, let x be the characteristic
node-vector:
Number of frustrated edges: Σ (xu − σuv xv)²
xu
=
+1/2 when u ∈ S
−1/2 when u ∈ T
uv∈E
∪
19. Algebraic Graph-theoretic Measures of Conflict 19Jérôme Kunegis
Frustration as Minimization
f is given by the solution to:
f = min Σ (xu − σuv xv)²
s.t. x ∈ {±1/2}V
Σ xu² = |V| / 4
⇔ ||x|| = |V| / 2
uv∈Ex
u
Relaxation
*
20. Algebraic Graph-theoretic Measures of Conflict 20Jérôme Kunegis
Using Matrices
The quadratic form can be expressed using matrices
Σ (xu − σuv xv)² = xT L x
where L ∈ ℝV × V is the matrix given by
Luv = −σuv when (uv) is an edge
Luu = d(u) is the degree of node u
L = D − A is the signed graph Laplacian
uv∈E
1
2
21. Algebraic Graph-theoretic Measures of Conflict 21Jérôme Kunegis
Minimizing Quadratic Forms
f* = min xT L x
s.t. ||x|| = |V| / 2
xT L x
xT x
x
f* = min
x
1
2
8
|V|
Rayleigh
quotient
min-max
theorem
f* = λmin
[L]8
|V|
f* = λmin
[L]
|V|
8
22. Algebraic Graph-theoretic Measures of Conflict 22Jérôme Kunegis
Relative Relaxed Frustration
Definition: Proportion of edges that have to be removed
to make the graph balanced
F* =
F* =
0 ≤ F* ≤ ≤ 1
λmin[L] ≤
f*
|E|
|V|
8 |E|
λmin
[L]
f
|E|
8 |E|
|V|
23. Algebraic Graph-theoretic Measures of Conflict 23Jérôme Kunegis
Properties of L (Unsigned Graphs)
L is positive-semidefinite (all λ[L] ≥ 0)
Multiplicity of λ = 0 equals number of connected
components
Smallest eigenvalue measures conflict
Second-smallest eigenvalue measures connectivity
(“algebraic connectivity”)
24. Algebraic Graph-theoretic Measures of Conflict 24Jérôme Kunegis
Properties of L (Signed Graphs)
L = Σ L(uv)
L(uv)
= [ ]when σuv = +1
L(uv)
= [ ]when σuv = –1
uv∈E
1 –1
–1 1
1 1
1 1
25. Algebraic Graph-theoretic Measures of Conflict 25Jérôme Kunegis
Minimal Eigenvalue of L (Signed Graphs)
L is positive-semidefinite (all eigenvalues ≥ 0)
L is positive-definite (all eigenvalues > 0) iff all
connected components are unlabanced
– Proof “ ”: by equivalence by all-positive graph⇒
– Proof “ ”: by contradiction (eigenvector would be all-zero)⇐
28. Algebraic Graph-theoretic Measures of Conflict 28Jérôme Kunegis
Computing λmin
[L]
Sparse LU decomposition + inverse power iteration:
O(|V|²) memory, but then very fast
% Matlab pseudocode
[ X Y ] = sparse_lu(L);
[ U D ] = eigs(@(x)(Y X x), k, 'sm');
32. Algebraic Graph-theoretic Measures of Conflict 32Jérôme Kunegis
Cross-Dataset Comparison of F*
For larger networks, a smaller
proportion of edges must be removed
to make it balanced
33. Algebraic Graph-theoretic Measures of Conflict 33Jérôme Kunegis
Merci
konect.uni-koblenz.de
Contact: Jérôme Kunegis
Université Koblenz-Landau
<kunegis@uni-koblenz.de> @kunegis
We want more datasets with negative edges, and
timestamps.
In particular: with changing and/or dissapearing edges