the math behind pagerank

The math behind PageRank

A detailed analysis of the mathematical aspects of PageRankComputational Mathematics class presentation

Ravi S SinhaLIT lab, UNT

Partial citations of references

• The Anatomy of a Large-Scale Hypertextual Web Search Engine Sergey Brin and Lawrence Page

• Inside PageRank Monica Bianchini, Marco Gori, and Franco Scarselli

• Deeper Inside PageRank Amy Langville and Carl Meyer

• Efficient Computation of PageRank Taher Haveliwala

• Topic Sensitive PageRank Taher Haveliwala

Overview of the talk

• Why PageRank• What is PageRank• How PageRank is used• Math• More math• Remaining math

Why PageRank• Need to build a better automatic search engine

Why?• Human maintained lists subjective and expensive to

build (non-automatic)• Automatic engines based on keyword matching do a

horrible job (just page content is not enough; cleverly placed words in a page can mislead search engines)

• Advertisers sometimes mislead search engines

• Solution: Google [modern day: much more than PageRank; getting smarter] Exact technology: not public domain Core technology: PageRank (utilizes link structure)

• Other uses Any problem that can be visualized as a graph

problem where the centrality of the vertices needs to be computed (NLP, etc.)

What is PageRank

• A way to find the most ‘important’ vertices in a graph

• PR(A) = (1-d) + d [ PR(T1) / C(T1) + … + PR(Tn) / C(Tn) ]

• Forms a probability distribution over the vertices [sum = 1]

• How does this relate to Web search? Vertices = pages Incoming edges = hyperlinks from other pages Outgoing edges = hyperlinks to other pages

Simple visualization: the simplest variant of PageRank in use [user behavior]

Random surfer

Damping factor

Only one incoming link, yet high PageRank

Lexical Substitution: A crash course

There are different types of managed care systems

Trivial for humans, not for machines

Math, statistics, linguistics wrapped within computer programs and algorithms

Information retrieval, machine translation, question answering, information security [information hiding in text]

PageRank in use: Lexical Substitution

Weights: word similarityDirected/ undirected: whole other realm

And now, the cool stuff

The math behind PageRank

• Intuitive correctness• Mathematical foundation• Stability• Complexity of computational scheme• Critical role of the parameters involved• The distribution of the page score• Role of dangling pages• How to promote certain vertices (Web pages)

Intuitive correctness

• Concept of ‘voting’ Related to citation in scientific literature More citations indicate great/ important piece of

work

• Random surfer / random walk• A page with many links to it must be important• A very important page must point to something

equally important

Mathematical foundation

• Most researchers: Markov chains Caveat: Only applicable in absence of dangling nodes

• Basic idea: authority of a Web page unrelated to its contents [comes from the link structure]

• Simple representation

• Vector representation

)1(][

dhxdx

ppq q

qp

NIdxWdx )1(

IN = [1, 1, 1 … 1]’

Transition matrix: ∑(each column) = 1 or 0

Mathematical foundation (2)

NId)()(txWd(t)x

11Google’s iterative version: converges to a stationary solution

Jacobi algorithm

NItN

txWdtx )1(1)1()(

Alternative computation

)1()1()1( txWdtxt

||x(t)||1 = 1; normalized

Web communities: Energy balance [measure of authority]

Ip

I

dpI

outI

inII

xE

EEEIE

*

More on energy

)( )()(

*1

*)1(1

*1 IOuti Idpi

iii

IIni

iiI

dpI

outI

inII

xddxf

ddxf

ddIE

EEEIE

Migration of scores across

graphLessons

Maximize energy

References from others

Minimize E(out)

Minimize E(dp)

Dangling pages, external links Maximize E(in)

inII EIE

Even more on energy [community promotion]

dpI

outI

inII EEEIE

1. Split same content into smaller vertices

2. Avoid dangling pages

3. Avoid many outgoing links

Page promotion

• Treat certain pages as communities• Bias certain pages by using a non-uniform

distribution in the vector IN

• Tinker with the connectivity [PageRank is proved to be affected by the regularity of the connection pattern]

NId)()(txWd(t)x

11Original •IN

•[1, 1, 1, …, 1]T

Biased •IN•[1, 1.5, 1.25, …, 1]T

Computation of PageRank

• PageRank can be computed on a graph changing over time Practical interest [Web is alive]

• An optimal algorithm exists for computing PageRank Practical applications: Search engines, PageRank on

billions of pages – efficiency! Ο(|Η| log 1/ε) NOT dependent on the connectivity or other

dimensions Ideal computation: stops when the ranking of vertices

between two computations does not change [converge]

The Markov model from the Web

• The PageRank vector can only exist if the Markov chain is irreducible

• By nature, the Web is non-bipartite, sparse, and produces a reducible Markov chain

• The Web hyperlinked matrix is forced to be Stochastic [non-negatives, all columns sum up to 1]

• Remove dangling nodes/ replace relevant rows/ columns with a small value, usually [1/n].eT

• Introduce personalization vector Primitive

• Non-negative• One positive element on the main diagonal• Irredicible

More on the Markov structure• A convex combination of the original stochastic

matrix and a stochastic perturbation matrix Produces a stochastic, irreducible matrix The PageRank vector is guaranteed to exist for this

matrix

• Every node directly connected to another node, all probabilities non zero Irreducible Markov chain, will converge

0 1/2

0 0

0 1/2

1/2 1/2

1/60 7/15

1/2 1/2

There’s more to PageRank

• Computation Power method

• Notoriously slow• Method of choice• Requires no computation of intermediate matrices• Converges quickly

Linear systems method

• The damping factor [usually 0.85] Greater value: more iterations required ‘Truer’ PageRanks

• Dangling pages• Storage issues

The end [for today]

Thanks for listening!