Navigating with Power Laws Aaron Clauset CAIDA / SFI Networks and - - PowerPoint PPT Presentation

navigating with power laws
SMART_READER_LITE
LIVE PREVIEW

Navigating with Power Laws Aaron Clauset CAIDA / SFI Networks and - - PowerPoint PPT Presentation

Navigating with Power Laws Aaron Clauset CAIDA / SFI Networks and Navigability Working Group 8 August 2008 Milgram Study (1967) A question of social connectedness 60 letters sent to Wichita, Kansas Destination: wife of divinity stud.,


slide-1
SLIDE 1

Navigating with Power Laws

Aaron Clauset CAIDA / SFI Networks and Navigability Working Group 8 August 2008

slide-2
SLIDE 2

Milgram Study (1967)

A question of social connectedness

  • 60 letters sent to Wichita, Kansas
  • Destination: wife of divinity stud., Cambridge, Ma.
  • Only 3 arrived
  • Subsequent studies: mean path length ~6

Discoveries

  • Surprisingly short paths; “small world” phenom.
  • Shorts paths are locally discoverable
  • S. Milgram, “The small world problem.” Psychology Today 2 (1967) 60--67.
slide-3
SLIDE 3

Watts-Strogatz Model (1998)

  • Modeled existence of short paths only
  • diameter log(n)

D.J. Watts and S.H. Strogatz, “Collective dynamics of small-world networks.” Nature 393 (1998) 440-442.

slide-4
SLIDE 4

Kleinberg Model (2000)

  • Model of navigability/search
  • Lattice + long range links
  • (Manhattan) distance metric
  • Local (greedy) navigation in ~ steps

log2(n)

  • J. Kleinberg. “The small-world phenomenon: an algorithmic perspective.”
  • Proc. 32nd ACM Symposium on Theory of Computing (2000) 163--170.

d(u, v) = |u − v|

slide-5
SLIDE 5

An Aside: Finite Size Effects

  • Simulate Kleinberg graphs with various
  • Measure mean routing time
  • Find severe finite size effects

e.g., for

  • Keep this in mind for later

α ∼ d T αTopt(n) = 1 − A log2(n) d = 1 αTopt = d

slide-6
SLIDE 6

Simulation Results

0.8 0.825 0.85 0.875 0.9 0.925 0.95 0.975 1 80 90 100 110 120 130 140

powerlaw exponent, α mean routing time, T αopt~0.900 n~2*105 αopt~0.911 n~5*105 αopt~0.921 n~1*106 αopt~0.928 n~2*106

slide-7
SLIDE 7

Convergence

10

1

10

2

10

3

10

4

10

5

10

6

10

7

10

−2

10

−1

10 network size, n 1 − αopt 1 − α(n) data 1 −2.8499log10

−2 (n)

slide-8
SLIDE 8

Not a new result

  • J. Kleinberg, “Navigation in a small world.” Nature 406 (2000) 845.

20,000 x 20,000 lattice

slide-9
SLIDE 9

Observation

  • Real networks often locally navigable

e.g., social network, world wide web Idea

  • Distributed changes to topology
  • Greedy changes to improve local navigability
  • web surfers on network of home pages
  • links changed based on speed of surfing

Origins of Navigability

slide-10
SLIDE 10
  • Same network as Kleinberg
  • Dynamic, greedy rewiring process
  • Global attractor for link-length distribution
  • Gives routing time

Clauset/Moore Model (2003)

u v a b

P() → −αrewired αrewired ∼ d T ∼ [log(n)]2

slide-11
SLIDE 11
  • 1. choose random pair (for )
  • 2. choose random tolerance on
  • 3. if routing time , become frustrated
  • 4. if frustrated, change random long-range link to have

length

Dynamics

[1, d(x, y)] (x, y) Tt Tt Tr ≥ Tt d = 1

slide-12
SLIDE 12

Simulation Results

Initial Conditions All self-loops, i.e., Stop Criteria When link-length distribution stabilizes

P() ∼ −∞

slide-13
SLIDE 13

Rewired Link-Lengths

1 4 19 84 369 1619 7098 31119 136423 598068 10 10

1

10

2

10

3

10

4

10

5

link−length (log−binned), l frequency f(l) = O(l−0.7789) n≈103 f(l) = O(l−0.8239) n≈104 f(l) = O(l−0.8454) n≈105 f(l) = O(l−0.8760) n≈106

slide-14
SLIDE 14

Routing Times

  • Measure mean routing times after stabilization
  • Fast routing times
  • Recall that (finite size effects)

T ∼ [log(n)]αTopt αTopt ∼ d

slide-15
SLIDE 15

Routing Times

10

1

10

2

10

3

10

4

10

5

10

6

10

7

25 50 75 100 125 150 network size, n mean routing time, T T(n) for α=d T(n) for αopt,αrewired T(n) =4.1log2(n)−7.39log(n) +6.08

slide-16
SLIDE 16

How long until navigable?

  • Let rewiring time be rewiring trials until
  • grows as a low-order polynomial

τ(n) Trewired ≤ 1.01 · Topt τ(n) τ(n) ∼ n1.77

slide-17
SLIDE 17

Time to Navigability

10

1

10

2

10

3

10

4

10

5

10

6

10

7

10 10

1

10

2

10

3

10

4

network size, n rewiring time, τ (trials per node) τ(n) data τ(n) =9+0.0304n0.77

slide-18
SLIDE 18

Global Attractor

Initial Condition

  • Link-length distribution
  • Measure as function of
  • Rewired distribution (eventually) independent of initial

condition αrewired α0

P() ∼ −α0

slide-19
SLIDE 19

Independence

0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 0.6 0.65 0.7 0.75 0.8 0.85 0.9 0.95 1 initial exponent, α0 rewired exponent, αrewired τ=75 τ=150 τ=300 τ=500

slide-20
SLIDE 20

Analytics

  • Distribution of tolerances
  • Otherwise, we don’t know much
  • If , then
  • What is ?

Tt P(Tt) = log n − log Tt n − 1 − log n α = d E[new length] = E[old length] P(frustrated | d(u, v))

slide-21
SLIDE 21
  • Navigability can come from distributed behavior
  • Natural/intuitive mechanism
  • Process is adaptive to changes in size, etc.
  • Analytics hard (full-history problems)
  • Power laws emerge spontaneously - why?
  • How does destination popularity effect rewired topol.
  • Preprint at cond-mat/0309415

Thoughts

slide-22
SLIDE 22