RTM: Laws and a Recursive Generator for Weighted Time-Evolving Graphs Leman Akoglu, Mary McGlohon, Christos Faloutsos Carnegie Mellon University School.

RTM: Laws and a Recursive Generator for Weighted Time-Evolving Graphs

Leman Akoglu, Mary McGlohon, Christos FaloutsosCarnegie Mellon University

School of Computer Science

1

Motivation Graphs are popular!

Social, communication,

network traffic, call graphs…

2

…and interesting surprising common

properties for static and un-weighted graphs

How about weighted graphs? …and their dynamic properties?

How can we model such graphs? for simulation studies, what-if scenarios, future prediction, sampling

Outline1. Motivation

2. Related Work - Patterns - Generators - Burstiness

3. Datasets

4. Laws and Observations

5. Proposed graph generator: RTM

6. (Sketch of proofs)

7. Experiments

8. Conclusion 3

Graph Patterns (I) Small diameter- 19 for the web [Albert and Barabási, 1999]- 5-6 for the Internet AS topology graph [Faloutsos, Faloutsos, Faloutsos, 1999]

Shrinking diameter

[Leskovec et al.‘05]

Power Laws

4

y(x) = Ax−γ, A>0, γ>0

Blog Network

time

diam

eter

Graph Patterns (II)

5

DBLP Keyword-to-Conference NetworkInter-domain Internet graph

Densification [Leskovec et al.‘05]

and Weight [McGlohon

et al.‘08] Power-laws Eigenvalues Power Law [Faloutsos et al.‘99]

Rank

Eig

enva

lue

|E|

|W|

|srcN|

|dstN|

Degree Power Law [Richardson and Domingos, ‘01]

In-degree

Cou

nt

Epinions who-trusts-whom graph

Graph Generators Erdős-Rényi (ER) model [Erdős, Rényi ‘60] Small-world model [Watts, Strogatz ‘98] Preferential Attachment [Barabási, Albert ‘99] Edge Copying models [Kumar et al.’99], [Kleinberg

et al.’99], Forest Fire model [Leskovec, Faloutsos ‘05] Kronecker graphs [Leskovec, Chakrabarti,

Kleinberg, Faloutsos ‘07] Optimization-based models [Carlson,Doyle,’00]

[Fabrikant et al. ’02]6

Edge and weight additions are bursty, and self-similar.

Entropy plots [Wang+’02] is a measure of burstiness.

Burstiness

Time

D W

eig

hts

Resolution

En

trop

y

From time series data, begin with resolution T/2. Record entropy HR.

Entropy plots

Time

D W

eig

hts

Resolution

En

trop

y

From time series data, begin with resolution T/2. Record entropy HR. Recursively take finer resolutions.

Entropy plots

Time

D W

eig

hts

Resolution

En

trop

y

Entropy Plots Self-similarity Linear plot

Resolution

En

trop

y

●

slope = 5.9

Entropy Plots Self-similarity Linear plot

Resolution

En

trop

y

time

Uniform: slope=1

slope = 5.9

Entropy Plots Self-similarity Linear plot

Resolution

En

trop

y

timetime

Uniform: slope=1 Point mass: slope=0

slope = 5.9

13 McGlohon, Akoglu, Faloutsos KDD08

Entropy Plots

Resolution

En

trop

yBursty:

0.2 < slope < 0.9

Self-similarity Linear plot

timetime

Uniform: slope=1 Point mass: slope=0

slope = 5.9

Outline1. Motivation

2. Related Work - Patterns

- Generators

3. Datasets

4. Laws and Observations

5. Proposed graph generator: RTM

6. Sketch of proofs

7. Experiments

8. Conclusion14

Datasets

15

Outline1. Motivation

2. Related Work - Patterns

- Generators

3. Datasets

4. Laws and Observations

5. Proposed graph generator: RTM

6. Sketch of proofs

7. Experiments

8. Conclusion16

Observation 1:λ1 Power Law (LPL)

Q: How does the principal eigenvalue λ1

change over time

A: λ1 (t) and the number of edges E(t) over time follow a power law with exponent less than 0.5,

especially after the ‘gelling point’.

17

λ1(t) E(t)∝ α,

α ≤ 0.5

λ1 Power Law (LPL) cont.

Theorem:For a connected, undirected graph G with N nodes and E edges, without self-loops and multiple edges;

18

DBLP Author-Conference network

Observation 2:λ1,w Power Law (LWPL)

Q: How does the weighted principal eigenvalue λ1,w change over time

A:

19

λ1,w(t) E(t)∝ β

DBLP Author-Conference network Network Traffic

Observation 3: Edge Weights PLQ: How does the weight of an

edge relate to “popularity” if its adjacent nodes

A: Weight of the link wi,j between two given nodes i and j in a given graph G has a power law relation with the weights wi and wj of the nodes;

20

FEC Committee-to-

Candidate network

Outline1. Motivation

2. Related Work - Patterns

- Generators

3. Laws and Observations

4. Proposed graph generator: RTM

5. Sketch of proofs

6. Experiments

7. Conclusion

21

Problem Definition Generate a sequence of realistic weighted

graphs that will obey all the patterns over time.

SUGP: static un-weighted graph properties small diameter power law degree distribution

SWGP: static weighted graph properties the edge weight power law (EWPL) the snapshot power law (SPL)

22

Problem Definition cont. DUGP: dynamic un-weighted graph properties

the densification power law (DPL) shrinking diameter bursty edge additions λ1 Power Law (LPL)

DWGP: dynamic weighted graph properties the weight power law (WPL) bursty weight additions λ1,w Power Law (LWPL)

23

One solution: Kronecker Product

24

Intuition: Self-similarity! Communities within

communities Recursion yields modular

network behavior

One solution: Kronecker Product

25

Recursive Tensor Product(RTM)

Use of tensors: 3rd mode is time Initial tensor I is a realistic graph itself

26

RTM of a (3x3x3) tensor by itself

RTM cont.

27

Outline1. Motivation

2. Related Work - Patterns

- Generators

3. Laws and Observations

4. Proposed graph generator: RTM

5. Sketch of proofs

6. Experiments

7. Conclusion

28

Experimental Results (I)

29

BLOG NETWORK

RTM MODEL

Experimental Results (II)

30RTM MODEL

BLOG NETWORK

Conclusion Largest (un)weighted principal eigenvalues are

power-law related to the number of edges in real graphs.

Weight of an edge is related to the total weights of its incident nodes.

Recursive Tensor Multiplication is a recursive method to generate weighted, time-evolving, self-similar, modular networks.

31

Future Directions Largest eigenvalues of the Laplacian matrices Second largest eigenvalue – related to global

connectivity – conductance – mixing rate of random walk on graph

Probabilistic version of RTM Fitting graphs

32

33

34