# DM561 - Linear Algebra with Applications

## Sheet 6

PageRank is the algorithm that Google originally used to determine the “importance” (or rank) of a web page.

The idea for PageRank is this: Define a Markov chain that describes the behavior of a random web-surfer. Consider the stationary distribution of this Markov chain. Define the weight of a page to be the probability of that page in the stationary distribution. Higher-probability pages are considered better. So when the user submits a query consisting of a set of words, present the web pages containing these words, in descending order of probability.

Conveniently, the PageRank vector (the stationary distribution) does not depend on any particular query, so it can be computed once and then used for all subsequent queries. (Of course, Google periodically recomputes it to take into account changes in the web.)

We start with a rudimentary Markov chain and we discover why it needs to be improved.

In each iteration, the random surfer selects an outgoing link from his current web page, and follows that link. We consider the small Web network of the picture, consisting of only four sites linked together as shown. Let $\delta{(j)}$ denote the number of links from page $j$ to other pages in the network. If $\delta{(j)}\neq 0$ and no other knowledge about the relevance of the pages is given, then it is reasonable to assume that the surfer follows links from the current web page with probability:

Write the transition matrix $A_1$ of the Markov chain and check that i ti indeed a stochastic matrix.

According to this Markov chain, how likely is that the random surfer is at each page after many iterations? What are the most likely pages?

Use the power method and carry out the operations in Python for the following cases:

• if it starts at page 6 and takes an even number of iterations

• if it starts at page 6 and takes an odd number of iterations

• if it starts at page 4 and takes an even number of iterations

• if it starts at page 4 and takes an odd number of iterations

You should find that the answer depends on where the surfer starts and how many steps he takes. The probabilities are almost the same except that the probabilities of nodes 2 and 3 are swapped.

From the point of view of computing definitive pageranks using the power method, there are two things wrong with this Markov chain:

i. There are multiple clusters in which the surfer gets stuck. One cluster is page 2 and page 3, and the other cluster is page 1.

ii. There is a part of the Markov chain that induces periodic behavior: once the surfer enters the cluster page 2 and page 3, the probability distribution changes in each iteration.

The first property implies that there are multiple stationary distributions. The second property means that the power method might not converge. We want a Markov chain with a unique stationary distribution so we can use the stationary distribution as an assignment of importance weights to web pages. We also want to be able to compute it with the power method. We apparently cannot work with the Markov chain in which the surfer simply chooses a random outgoing link in each step.

Consider then an alternative very simple Markov chain: the surfer jumps from whatever page he’s on to a page chosen uniformly at random.

Write the transition matrix $A_2$ for this Markov chain. Is there a unique stationary distribution? Which one?

As you may have discovered the latter Markov chain does not in any way reflect the structure of the Web network. Using the stationary distribution to assign weights would provide no information at all.

Instead, we will use a mixture of these two Markov chains. That is, we will use the Markov chain whose transition matrix is

Show that the matrix $A$ is a stochastic matrix, that is, the sum of the entries in all columns is 1.

The Markov chain corresponding to the matrix $A$ describes a surfer obeying the following rule:

• With probability 0.85, select one of the links from the current web page, and follow it.

• With probability 0.15, jump to a web page chosen uniformly at random. (This is called teleporting in the context of PageRank.)

You can think of the second item as modeling the fact that sometimes the surfer gets bored with where he/she is. However, it plays a mathematically important role. The matrix $A$ is a positive matrix (every entry is positive). A theorem ensures that there is a unique stationary distribution, and that the power method will converge to it.

Calculate how likely it is that the random surfer is at each page after many iterations and rank consequently the pages. You can use the theory of diagonalization or the power method and carry out the calculation with Python.

The file top250movies.txt contains data from the 250 top-rated movies according to IMDb (scraped with imdbpy). Each line in the file lists a movie title and its cast as title/actor1/actor2/..., with the actors listed mostly in billing order (stars first), though some casts are listed alphabetically or in order of appearance.

Create a networkx nx.DiGraph object with a node for each actor in the file. The weight from actor a to actor b should be the number of times that actor a and b were in a movie together but actor b was listed first. That is, edges point to higher-billed actors.

• Compute the PageRank values of the actors.

• Write a function to rank them

• Return the list of ranked actors.

(Hint: Consider using itertools.combinations() while constructing the graph. Also, use encoding="utf-8" as an argument to open() to read the file, since several actors and actresses have nonstandard characters in their names such as ø and æ.)

With $\epsilon = 0.7$, the top three (most visible) actors should be Leonardo DiCaprio, Robert De Niro, and Tom Hanks, in that order.