BIG DATA ANALYTICS

CS-411

Dr. Bharat Singh

Department of Computer Science and Engineering

Indian Institute of Information Technology, Ranchi

 11/12/24
Page Rank Algorithm
“A page is important if important pages link to it.”

2
 PAGE RANK ALGORITHM

11/12/24
 The PageRank algorithm or Google algorithm was introduced by Lary Page, one of the founders of
Google.
 It was first used to rank web pages in the Google search engine.

 Nowadays, it is more and more used in many different fields, for example in ranking users in social

media etc…
 What is fascinating with the PageRank algorithm is how to start from a complex problem and end up

with a very simple solution.

3
 PAGE RANK ALGORITHM

11/12/24
The web can be represented like a directed
graph where nodes represent the web
pages and edges form links between them.
Typically, if a node (web page) i is linked to
a node j, it means that i refers to j.

4
 PAGE RANK ALGORITHM
 We have to define what is the importance of a web page.
 As a first approach,

11/12/24
 it is the total number of web pages that refer to it.
 If we stop to this criteria, the importance of these web pages that refer to it is not taken into account.
 In other words, an important web page and a less important one has the same weight.
 Another approach is
 a web page spread its importance equally to all web pages it links to.
 By doing that, we can then define the score of a node j as follows:

 where rᵢ is the score of the node i and dᵢ its out-degree.

5
 PAGE RANK ALGORITHM

11/12/24
From the graph, we can write this linear system:

we can solve by using Gaussian elimination.

6
 PAGE RANK ALGORITHM

11/12/24
 But this solution is limited for small graphs.
 Indeed, as this kind of graphs are sparse and Gauss elimination

modifies the matrix when performing its operations, we lose the

sparsity of the matrix and it would take more memory space.
 In the worst case, the matrix can no longer be stored.

7
 PAGE RANK ALGORITHM

11/12/24
 Markov Chain and PageRank
 Graph can be seen as a Markov chain
with the following transition matrix:

8
 PAGE RANK ALGORITHM

11/12/24
 P transpose is row stochastic which is a condition to apply Markov chain theorems.
 For the initial distribution, let’s consider that it is equal to :

 where n is the total number of nodes.

 This means that the random walker will choose randomly the initial node from where it can reach all other

nodes.


9
 PAGE RANK ALGORITHM
 At every step, the random walker will jump to another node according to the transition matrix. the
probability distribution is then computed for every step.

11/12/24
 This distribution tells us where the random walker is likely to be after a certain number of steps.

 The probability distribution is computed using the following equation:

 A stationary distribution of a Markov chain is a probability distribution π with π = Pπ.

 This means that the distribution will not change after one step. It is important to note that not all Markov

chains admit a stationary distribution

10
 PAGE RANK ALGORITHM

11/12/24
 All we have to do is solving this equation:

 We notice that π is an eigenvector of the matrix P with the eigenvalue 1.

 Instead of computing all eigenvectors of P and select the one which corresponds to the

eigenvalue 1, we use the Frobenius-Perron theorem.

11
 PAGE RANK ALGORITHM

11/12/24
Frobenius-Perron theorem:
 If a matrix A is a square and positive matrix (all its entries are positive),
then it has a positive eigenvalue r, such as |λ| < r, where λ is an
eigenvalue of A. The eigenvector v of A with eigenvalue r is positive and is
the unique positive eigenvector.
 To compute π, we use the power method iteration which is an iterative method to compute the
dominant eigenvector of a given matrix A.

12
 PAGE RANK ALGORITHM
Teleportation and Damping Factor
 In the web graph, we can find a web page i which refers

11/12/24
only to web page j and j refers only to i. This is what we
call spider trap problem.
 We can also find a web page which has no outlink. It is

commonly named Dead end.



13
 PAGE RANK ALGORITHM

11/12/24
 Spider Trap, when the random walker reaches the node 1 in the above example, he can
only jump to node 2 and from node 2, he can only reach node 1, and so on. The
importance of all other nodes will be taken by nodes 1 and 2. In the above example, the
probability distribution will converge to π = (0, 0.5, 0.5, 0). This is not the desired result.
 Dead Ends, when the walker arrives at node 2, it can’t reach any other node because it

has no outlink. The algorithm cannot converge.

 To get over these two problems, we introduce the notion
of teleportation.

14
 PAGE RANK ALGORITHM

Teleportation

11/12/24
 Teleportation consists of connecting each node of the graph to all other nodes.
 The graph will be then complete.
 The idea is with a certain probability β, the random walker will jump to another node according
to the transition matrix P and with a probability (1-β)/n, it will jump randomly to any node in the
graph. We get then the new transition matrix R:

where v is a vector of ones, and e a vector of 1/n. β is commonly defined as the damping factor.

15
 PAGE RANK ALGORITHM
TELEPORTATION
 By applying teleportation in our example, we get the following new transition matrix:

11/12/24
 The matrix R has the same properties than P which means that it admits a stationary distribution,
so we can use all the theorems we saw previously.

16
 PAGE RANK ALGORITHM

11/12/24


17
 Thank You!!
Dr. Bharat Singh
Email id—
[email protected]
Mobile No–
8707223885