22
Power Law and Its Generative Models Bo Young Kim 2010-03-16

Power Law and Its Generative Models

  • Upload
    caron

  • View
    94

  • Download
    4

Embed Size (px)

DESCRIPTION

Power Law and Its Generative Models. Bo Young Kim 2010-03-16. Contents. Recall The Definition of Power Law Recall Some Properties of Power Law Generative Models for Power Law - Power Laws via Preferential Attachment - Power Laws via Multiplicative Processes. - PowerPoint PPT Presentation

Citation preview

Page 1: Power Law  and Its Generative Models

Power Law and Its Generative Models

Bo Young Kim2010-03-16

Page 2: Power Law  and Its Generative Models

Applied Algorithm Lab. 2

Contents1. Recall The Definition of Power Law2. Recall Some Properties of Power Law3. Generative Models for Power Law

- Power Laws via Preferential Attachment- Power Laws via Multiplicative Processes

Page 3: Power Law  and Its Generative Models

Applied Algorithm Lab. 3

1. Recall The Definition of Power Law2. Recall Some Properties of Power Law3. Generative Models for Power Law

- Power Laws via Preferential Attachment- Power Laws via Multiplicative Processes

Page 4: Power Law  and Its Generative Models

Applied Algorithm Lab. 4

1. Recall The Definition of Power Law

• X: a nonnegative random variable

• Def Power LawX is said to have a power law distribution if Pr[X≥x]~cx-α for constants c>0, α>0

• Def f(x)~g(x) ⇔ limx f(x)/g(x) = 1

• What does this mean?In a power law distribution, asymptotically the tails fall according to the power α. (heavier tail than exponential distribution)

Page 5: Power Law  and Its Generative Models

Applied Algorithm Lab. 5

1. Recall The Definition of Power Law2. Recall Some Properties of Power Law3. Generative Models for Power Law

- Power Laws via Preferential Attachment- Power Laws via Multiplicative Processes

Page 6: Power Law  and Its Generative Models

Applied Algorithm Lab. 6

2. Recall Some Properties of Power Law

• E.g. The Pareto distributionPr[X≥x]=(x/k)-α

ln(Pr[X≥x])=-α(ln(x)-ln(k))

* Linear Log-log plot (complementary cumulative distribution function)

- X has a power law distribution- Then a log-log plot behavior is a straight line. (asymptotic sense)

Page 7: Power Law  and Its Generative Models

Applied Algorithm Lab. 7

2. Recall Some Properties of Power Law

“Scale Invariance”- Let f(x) := P[X≥x]- f(x) ~ cx-α

- f(kx) ~ c(kx) -α = k-α(cx-α ) = k’f(x) ∝ f(x) (k’=k-α)

- Scaling by a constant simply multiplies the original power law relation by the constant k’.- If we change the measurement unit(=scale), it re-tains the same power law form w/ the same expo-nent. We cannot decide what scale we’re observing. (like Fractals)

Page 8: Power Law  and Its Generative Models

Applied Algorithm Lab. 8

2. Recall Some Properties of Power Law

• Web follows power law. [4]• Recall (Rank exponent)

- dv: outdegree of a node v - rv: the rank of a node v

dv =k*rvR (R,k: constant)

• Designing random graph models that yield Web-like graphs?

• i.e. that yields power law distributions for the in-degree and outdegree?

Page 9: Power Law  and Its Generative Models

Applied Algorithm Lab. 9

1. Recall The Definition of Power Law2. Recall Some Properties of Power Law3. Generative Models for Power Law

- Power Laws via Preferential Attachment- Power Laws via Multiplicative Processes

Page 10: Power Law  and Its Generative Models

Applied Algorithm Lab. 10

Generative Models for Power Law- Power Laws via Preferential Attachment

• Def Preferential Attachment Process (=Yule Process)Any process s.t. some quantity (some form of wealth) is distributed among a number of individ-uals according to how much they already have, so that those who are already wealthy receive more than those who are not.

• ”The rich get richer”

Page 11: Power Law  and Its Generative Models

Applied Algorithm Lab. 11

• The Chinese Restaurant Process- A Chinese restaurant has infinitely many tables- Each table can seat infinitely many customers- At each time step, customer Xt comes into the restau-rant.When Xt+1 comes into here…

(CRP1) Sits at an already occupied table k w/ prob. Nk/(t+α)(Nk : # of customers at table k ΣkNk=t)(CRP2)or, sits at the next unoccupied table w/ prob. α/(t+α)

Generative Models for Power Law- Power Laws via Preferential Attachment

Page 12: Power Law  and Its Generative Models

Applied Algorithm Lab. 12

When Xt+1 comes into here…

(CRP1) Sits at an already occupied table k w/ prob. Nk/(t+α)(Nk : # of customers at table k ΣkNk=t)(CRP2)or, sits at the next unoccupied table w/ prob. α/(t+α)

Generative Models for Power Law- Power Laws via Preferential Attachment

Page 13: Power Law  and Its Generative Models

Applied Algorithm Lab. 13

• CPR rule: Next customer sits at a table w/ prob. Proportional to # of customers already sitting at it(and sits at new table w/ prob. Proportional to α) Customers tend to sit at most popular tables Most popular tables attract the most new cus-tomers, and become even more popular

• The concentration parameter α: how likely cus-tomer is to sit at a fresh table

Generative Models for Power Law- Power Laws via Preferential Attachment

Page 14: Power Law  and Its Generative Models

Applied Algorithm Lab. 14

• Generating Power law distribution via Preference At-tachment (Most models are variations of this form)

• Let’s say “Web Page Process”• Start w/ a single page• This single page has a link to itself• At each time step, a new page appears, w/ outdegree 1

Generative Models for Power Law- Power Laws via Preferential Attachment

(WPP1) The link of new page points to a page chosen u.a.r. w/ prob. α<1 (WPP2) The link of new page points to page chosen proportionally to the indegree of the page w/ prob. 1- α

Page 15: Power Law  and Its Generative Models

Applied Algorithm Lab. 15

• Xj(t): # of pages w/ indegree j when ∃ t pages in the system

• Pr[Xj increase] = αXj-1/t+(1-α)(j-1)Xj-1/t• Pr[Xj decrease] = αXj/t+(1-α)jXj/t

Generative Models for Power Law- Power Laws via Preferential Attachment

(WPP1) The link of new page points to a page chosen u.a.r. w/ prob. α<1 (WPP2) The link of new page points to page chosen proportionally to the indegree of the page w/ prob. 1- α

Page 16: Power Law  and Its Generative Models

Applied Algorithm Lab. 16

• Pr[Xj increase] = αXj-1/t+(1-α)(j-1)Xj-1/t• Pr[Xj decrease] = αXj/t+(1-α)jXj/t

dXj/dt = {α(Xj-1-Xj)+(1-α)((j-1)Xj-1-jXj-1)}/t

• Intuitively appealing, BUT how continuous DE de-scribes a discrete process?

This can be justified formally using martingales [Kumar et al 00] & theoretical frameworks of Kurtz, Wormald [Drinea et al. 00, Kurtz 81, Wormald 95].

Generative Models for Power Law- Power Laws via Preferential Attachment

Page 17: Power Law  and Its Generative Models

Applied Algorithm Lab. 17

• dX0/dt=1-αX0/t

• Suppose in the steady state limit: Xj(t)=cj*t (portion cj) c0 =dX0/dt=1-αX0/t=1-αc0 ⇔ c0 = 1/(α+1)

• Substitute this assumption for dXj/dt = {α(Xj-1-Xj)+(1-α)((j-1)Xj-1-jXj-1)}/t cj(1+α+j(1-α))=cj-1(α+(j-1)(1-α)) We can determine cj exactly.

• Focusing on the asymptotic, for large jcj/cj-1=1-(2-α)/(1+α+j(1-α))~1-{(2-α)/(1-α)}*(1/j)

Generative Models for Power Law- Power Laws via Preferential Attachment

Page 18: Power Law  and Its Generative Models

Applied Algorithm Lab. 18

• We have cj~cj^(- ) for some constant c, giving a power law.

• Note cj~cj^(- ) implies

WTS: Σj≥kcj behave the tail of power law distribution(Proof)

For some constant c’. So, we’re done.

Generative Models for Power Law- Power Laws via Preferential Attachment

12

12 )1)(

12(1~)1(~ 1

2

1 jjj

cc

j

j

1

1

'~~~ 12

12

kcdjcjcjckjkjkj

j

Page 19: Power Law  and Its Generative Models

Applied Algorithm Lab. 19

1. Recall The Definition of Power Law2. Recall Some Properties of Power Law3. Generative Models for Power Law

- Power Laws via Preferential Attachment- Power Laws via Multiplicative Processes

Page 20: Power Law  and Its Generative Models

Applied Algorithm Lab. 20

• Pareto: income distribution obeys power law• [Champernowne 53] offered an explanation for this be-

havior.• Partition income in the following manner:• 1st range: between m and γm for some γ>1• 2nd range: between γm and γ2m …• persons in class j: their income is between γj-1m and γjm • Pij: prob. of a person moving from class i to class j• At each time step, Pij depends only on the value (j-i).

Under this assumption, Pareto distribution can be ob-tained.

Generative Models for Power Law- Power Laws via Multiplicative Processes

Page 21: Power Law  and Its Generative Models

Applied Algorithm Lab. 21

• E.g. γ=2, Pij=2/3 if j-i=-1 Pij=1/3 if j-i=1

• Special case: i=1 P11=2/3

• The equilibrium property of being in class k: 1/2k

X: a person’s income Pr[X≥2k-1m]=1/2k-1

Pr[X ≥ x]=m/x for x= 2k-1m

This is a power law distribution.

Generative Models for Power Law- Power Laws via Multiplicative Processes

Page 22: Power Law  and Its Generative Models

Applied Algorithm Lab. 22

References[1] M. Mitzenmacher, A Brief History of Generative Models for Power Law and

Lognormal Distributions, Internet Mathematics, vol 1, No. 2, pp. 226-251, 2004. [2] Mark Johnson, Chinese Restaurant Processes(CG168 notes), cog.brown.edu/

~mj/classes/cg168/.../ChineseRestaurants.pdf[3] The lecture notes of C. Faloutsos, Carnegie Mellon University, 15-826 Multimedia

Databases and Data Mining, Spring 2008 http://www.cs.cmu.edu/~christos/courses/826.S08/FOILS-pdf/195_powerLaws.pdf

[4] Bruno Bassetti, Mina Zarei, Marco Cosentino Lagomarsino, and Ginestra Bianconi., Statistical mechanics of the “Chinese restaurant process”: Lack of self-av-eraging, anomalous finite-size effects, and condensation, Phys. Rev. E 80, 066118 (2009) [4 pages]

[5] http://en.wikipedia.org/wiki/Power_law, http://en.wikipedia.org/wiki/Chinese_restaurant_process, http://en.wikipedia.org/wiki/Preferential_attachment