Digital Garden
Machine Learning
Link Analysis

Link Analysis

Page Rank

very earily in times of internet how to organize web, first try was human curated web directoreis like yahoo, dmoz, look smart.

second try was web search, google. Wanted to find relevant docs from a small and trsuted set but web is huge full of untrusted documents and random pages.

Challanges wo to trust? The idea for this is that trustowrthy pages point to each other. What is best answer for query? The idea for this is that pages that know about a subject are pointing to many other pages

Flow Formulation

Idea is that links are votes, a page is more important if it has more links, do we look at in or out links? are all links equal?, links from important pages should count more => recursive question.

Simple recursive formulation is that each links vote is proportional to the importance of its source page. so if page j has important rj with n out link each link should get rj/n votes, js own importance rj is the sum of the votes of its inlinks. A vote from an important page is worth more this also fullfills the requirement of a page is important if it is pointed to by orther importange pages. this leads to n equations for n nodes and no constants so there is no unique solution. so to get a unique solution we add a constraint that all page ranks summed together equal 1. We can solve this by using gaussian elimation but this becomes very quickly very hard for large web graphs so we need a new formulation.

matrix formulation, instead of the equations we write a column stochastic adjacency matrix, meaning each column sums up to 1. so if page i has di out links and there is a connection from i to j then Mji is 1/di otherwise 0.

we then have the rank vector r which contains all the page rankings. so ri is the score for page i and all score summed = 1 as before. we can then write the flow equaiton as r=M*r.

In other words the rank vector is an eigenvector of the adjacency matrix m with the eigenvalue 1. because the matrix is column stochstic the largest eigenvalue is 1.

Now to solve these flow equations we can use the power iteration method. Suppose there are N pages we initialize rank vector r0 with all values 1/N. we then iterate rt+1 = M*rt and stop when we no longer change the values much, i.e we converge. rt+1 -rt < epsilon. This is a method to find the dominant eigenvector i.e the vector corresponding to the largest eigenvalue which because it is column stochastic is 1.

r1 = M * r0 r2 = Mr1 = M(Mr1) = M^2 * r0 etc.

We can also interpret this using probailities on a random walk. the pagerank vector gives the stationary distribution for a random walk on a graph i.e

Google Formulation

There is a problem with the above definition as it might not always converge. Google defined two types of problems that can occur in a network which stops the power iteration from converging.

dead ends i.e pages with no out links so a random walk can not leave so eventually importance leaks out to 0.

Spider traps all out links are within a group so random walk get stuck in a trap. these traps then slowly absorob all the importance.

The soultion to both of these issues is teleports. To solve spider traps at each step the random walker/surfer has two options with a prob of beta it will follow a link at random with 1-beta it will teleport to some random page, common values for beta are 0.8 or 0.9.

To solve the issue of dead ends we just adjust the adjacancy matrix by keeping the colum stochastic properity, so if a node has no out links it just becomes a unfirom distribution to all nodes.

This slightly changes the page rank quation as now there is the additional probability of a teleport. this leads to the google matrix which can still be solved with the power iteration method.

But what if we have 1 billion pages we can not really store this in main memory so we need to make use of the spasity of the matrix. we can do this by just storing the degree for each source node and its destination nodes. we can then store rold and M on the disk and rnew in memory and just calcualte row by row so we do not need to load the entire matrix into memory.

But what if we can not even fit the entire rnew in memory, then we can use the so called block based update algorithm where we then just do block by block.

In the above solution we are still scanning m and rold once for each block tho so we can do even better. which leads to block stripe update algorithm. Each strip contains only destination nodes corresponding to the blocks of rnew.

Page rank does however have its limitaitons. It only measures generic popularity of a page. The solution to this is topic specific page rank. It only uses a single measure of importance. Other models of importance are used in the hub and authorities extension. It is susceptible to link spam, i.e artificial link topographies to boost page rank to solve the there is the trust rank.

Topic specific page rank is useful if we do not just want a generic popularity but a popularity within a topic. this allows search quieries to be answwered based on intersts of the users for example the query jaguar is it the animal or car?

The idea here is a bias random walk so when teleporting instead of going to any page uniformly we only teleport to a topic-specific set. this gives pages closer to these topic specific pages a higher rank. to make this work all we need to do is update the teleportation part and give the ones with topic specific higher probability to be teleported to.

proximity on graphs i.e the relevance closeness or similarity of pages. shortest path is not good because slow and also does not take into account hubs. network flow is also not good as it does not punish long paths? dont know netowork flow lol.

Random Walk with Restart. makes no sense

What is web spam? any deliberate action to boost web pages position. spam are web pages that are the result of spamming. Constant war, first search engines considered number of times query word appeared, prominece of word position. Add keywords hidden thousand of times, or links with high topic ranking. Googles solution is to look at what others say about u for example looking at anchor text that is linked to ur page.

spam farming. Inaccessible pages, accessible pages, owned pages. To combat spam the trsut rank was developed. which is topic specific page rank but with a teleport set of trusted pages like .edu or .gov pages.

We create a seed set by hand which contains good trusted pages. We can then use this as the teleport set. Trust attenuation is the degreee of trust which is spread across the graph. trust splitting across out links

spam mass = what fraction of a pages rank comes from spam pages. page rank p rp, rp+ page rank with trusted pages only in teleport page rank from spam pages is then the difference and the spam mass is the proportion.

HITS, hypertextinduced topic selection is a meassure of importance. we have hubs and authorities. Hubs are course bulleting, authoritizes hold usefull information. Authority each page start with hub score 1 authorities collect there votes, hubs collect authority scores. so it doesn't just infetly increase it is normalized that all hubs equal 1.