In this thesis, I study the spectral characteristics of large dynamic networks and formulate the spectral evolution model. The spectral
evolution model applies to networks that evolve over time, and describes their spectral decompositions such as the eigenvalue and singular value
decomposition.
The spectral evolution model states that over time, the eigenvalues of a
network change while its eigenvectors stay approximately constant.
On the Spectral Evolution of Large Networks (PhD Thesis by Jérôme Kunegis)
1. Web Science & Technologies
University of Koblenz ▪ Landau, Germany
On the Spectral Evolution
of Large Networks
Jérôme Kunegis
2. Networks
…are everywhere
ip
r sh
tho
Au
ip
dsh
Fr ien
t
Trus
o n
ic ati
n
mu e
Co
m e nc
c
ti on c urr
ra c -oc
I nte Co
Jérôme Kunegis
kunegis@uni-koblenz.de 2 / 29
3. Social Network
Person Friendship
Jérôme Kunegis
kunegis@uni-koblenz.de 3 / 29
7. Recommender Systems
:-(
me
Predict who I will add as friend next
Facebook's algorithm: find friends-of-friends
→ Problem: Rest of the network is ignored!
Jérôme Kunegis
kunegis@uni-koblenz.de 7 / 29
8. Outline
1. Algebraic Link Prediction
2. Spectral Transformations
3. Learning Link Prediction
Take into account the whole network
Jérôme Kunegis
kunegis@uni-koblenz.de 8 / 29
9. Adjacency Matrix
3
1 2 4 5 6
1 2 3 4 5 6
Aij = 1 when i and j are connected 1 0 1 0 0 0 0
Aij = 0 when i and j are not connected
2 1 0 1 1 0 0
3 0 1 0 1 0 0
A=
4 0 1 1 0 1 0
A is square and symmetric
5 0 0 0 1 0 1
6 0 0 0 0 1 0
Jérôme Kunegis
kunegis@uni-koblenz.de 9 / 29
10. Baseline: Friend of a Friend Model
Count the number of ways a person can be found as
the friend of a friend
Matrix product AA = A2
2 3
0 1 0 0 0 0 1 0 1 1 0 0
1 0 1 1 0 0 0 3 1 1 1 0
A 2
=
0
0
1
1
0
1
1
0
0
1
0
0
=
1
1
1
1
2
1
1
3
1
0
0
1
0 0 0 1 0 1 0 1 1 0 2 0 1 2 4
0 0 0 0 1 0 0 0 0 1 0 1
Jérôme Kunegis
kunegis@uni-koblenz.de 10 / 29
12. Eigenvalue Decomposition
A = UΛUT
where
U are the eigenvectors U TU = I
Λ are the eigenvalues Λij = 0 when i ≠ j
Jérôme Kunegis
kunegis@uni-koblenz.de 12 / 29
13. Computing A3
Use the eigenvalue decomposition A = UΛUT
A3 = UΛUT UΛUT UΛUT = UΛ3UT
Exploit U and Λ:
U TU = I because U contains eigenvectors
(Λ ) = Λiik because Λ contains eigenvalues
k
ii
Result: Just cube all eigenvalues!
Jérôme Kunegis
kunegis@uni-koblenz.de 13 / 29
15. Spectral Transformations
A2 = UΛ2UT Friend of a friend
A = UΛ UT
3 3
Friend of a friend of a friend
exp(A) = Uexp(Λ)UT Matrix exponential
…are link prediction functions!
Jérôme Kunegis
kunegis@uni-koblenz.de 15 / 29
16. Outline
1. Algebraic Link Prediction
2. Spectral Transformations
3. Learning Link Prediction
Why does it work?
Jérôme Kunegis
kunegis@uni-koblenz.de 16 / 29
17. Looking at Real Facebook Data
Dataset: Facebook New Orleans
(Viswanath et al. 2009)
63,731 persons
1,545,686 friendship links with creation dates
Adjacency matrix At at time t (t = 1 . . . 75)
Compute all eigenvalue decompositions At = UtΛtUtT
Jérôme Kunegis
kunegis@uni-koblenz.de 17 / 29
20. Eigenvector Permutation
Time split: old edges A = UΛUT
new edges B = VDVT
Eigenvectors permute
|U•i∙ V•j|
Jérôme Kunegis
kunegis@uni-koblenz.de 20 / 29
21. Outline
1. Algebraic Link Prediction
2. Spectral Transformations
3. Learning Link Prediction
a) Learning by Extrapolation
b) Learning by Curve Fitting
What spectral transformation is best?
Jérôme Kunegis
kunegis@uni-koblenz.de 21 / 29
22. a) Learning by Extrapolation CIKM 2010
Extrapolate the growth of the spectrum
Good when growth
is irregular
Potential problem:
overfitting
Jérôme Kunegis
kunegis@uni-koblenz.de 22 / 29
23. b) Learning by Curve Fitting
f
A B
UΛUT B
Λ UTBU
f
Diagonal
Jérôme Kunegis
kunegis@uni-koblenz.de 23 / 29
25. Polynomial Curve Fitting
Fit a polynomial a + bx + cx2 + dx3 + ex4
Jérôme Kunegis
kunegis@uni-koblenz.de 25 / 29
26. Other Curves ICML 2009
Friend of a Friend
a + bx + cx2
Polynomial
a + bx + cx2 + dx3 + ex4
Nonnegative polynomial
a + . . . + hx7 a, . . ., h ≥ 0
Matrix exponential
b exp(ax)
Neumann kernel
b / (1 − ax)
Rank reduction
ax if |x| ≥ x0, 0 otherwise
Jérôme Kunegis
kunegis@uni-koblenz.de 26 / 29
27. Evaluation Methodology
All edges E
Training set Ea ∪ Eb
˙ Apply Test set Ec
Source set Ea Learn Target set Eb
Edge creation time
3-way split of edge set by edge creation time
Jérôme Kunegis
kunegis@uni-koblenz.de 27 / 29
28. Experiments
Precision of link prediction (1 = perfect)
All datasets available at konect.uni-koblenz.de
Jérôme Kunegis
kunegis@uni-koblenz.de 28 / 29
29. Conclusion
●
Observation
●
Eigenvalue change, eigenvectors are constant
●
Why?
●
Graph kernels, triangle closing, the sum-over-paths model,
rank reduction, etc.
●
Application to recommender systems
●
By learning the spectral transformation for a given dataset
ACKNOWLEDGMENTS →
Thank You!
Jérôme Kunegis
kunegis@uni-koblenz.de 29 / 29
30. Selected Publications
The Slashdot Zoo: Mining a social network with negative edges
J. Kunegis, A. Lommatzsch and C. Bauckhage
In Proc. World Wide Web Conf., pp. 741–750, 2009.
Learning spectral graph transformations for link prediction
J. Kunegis and A. Lommatzsch
In Proc. Int. Conf. on Machine Learning, pp. 561–568, 2009.
Spectral analysis of signed graphs for clustering, prediction and
visualization
J. Kunegis, S. Schmidt, A. Lommatzsch and J. Lerner
In Proc. SIAM Int. Conf. on Data Mining, pp. 559–570, 2010.
Network growth and the spectral evolution model
J. Kunegis, D. Fay and C. Bauckhage
In Proc. Conf. on Information and Knowledge Management,
pp. 739–748, 2010.
Jérôme Kunegis
kunegis@uni-koblenz.de 30 / 29
31. References
B. Viswanath, A. Mislove, M. Cha, K. P. Gummadi, On
the evolution of user interaction in Facebook. In Proc.
Workshop on Online Social Networks, pp. 37–42, 2009.
Jérôme Kunegis
kunegis@uni-koblenz.de 31 / 29