Slashdot Mirror
The Man Behind Google's Ranking Algorithm
nbauman writes "New York Times interview with Amit Singhal, who is in charge of Google's ranking algorithm. They use 200 "signals" and "classifiers," of which PageRank is only one. "Freshness" defines how many recently changed pages appear in a result. They assumed old pages were better, but when they first introduced Google Finance, the algorithm couldn't find it because it was too new. Some topics are "hot". "When there is a blackout in New York, the first articles appear in 15 minutes; we get queries in two seconds," said Singhal. Classifiers infer information about the type of search, whether it is a product to buy, a place, company or person. One classifier identifies people who aren't famous. Another identifies brand names. A final check encourages "diversity" in the results, for example, a manufacturer's page, a blog review, and a comparison shopping site."
Pigeon Rank?
... is not to be confused with Amit Singh, who also works at Google and has authored an excellent book on Mac OS X Mac OS X Internals.
No, but I DO see the difference between 'appleS' and 'apple', just as the text you're quoting mentions.
> They use 200 "signals" and "classifiers," of which PageRank is only one.
How many did they expect PageRank to be? In the words of someone immortal, "There can be only one.".
Max.
Oh yeah. Woops. That isn't as interesting :-)
In Soviet Russia, they shoot idiots why don't realize this joke is dead.
My ongoing gripe with Google is the number of times when the first page is filled with shopping sites, "review" pages, and click through pages that exist only to grab you onto the way to where you really want to go.
I would love a switch, or even a subscription, that would allow me to filter these usually useless types of pages and instead show me pages with real content.
Three Squirrels
Pagerank is the source of all wisdom in google... but there is so much more... Like string searching & matching algos, file searching.. you name it.. Just the other day I was searching for books about Google's algorithms... I found zero interesting stuff.. They keep their algorithms secret and out of the public domain... (like they should..). we praise Pagerank, but if we knew what other stuff is there, we would all be members of Church of Google (http://www.thechurchofgoogle.org/) :P
God had a 7 day deadline... So he made the world in LISP
Do not try to read the dupe, thats impossible. Instead, only try to realize the truth
What truth?
There is no dupe
One of the most annoying things about google for me is how it interprets queries with strange characters common to almost all programming languages. A google search for "ruby <<" returns no results related to the ruby append operator. A Simple search for "<<", by itself returns ZERO results.
This could allow for a better search result when using for example "APPLE NEAR MACINTOSH" or "APPLE NEAR BEATLES"
Ho hum... Times changes and not always for the better...
If builders built buildings the way programmers wrote programs, then the first woodpecker would destroy civilization.
Actually, using -site:.co.uk would yield much better results. Since he will then get everything except .co.uk instead of just .com
Get a web developer
Does the algorithm account for the toilet seat's positon?
The AACS key is NOT 0xF606EEFD628B1CA427BEA93A9CA9773F
One interesting thing about the article was the down-to-earth lack of abstraction in the problems described, such as the teak patio palo alto problem. Other search engines brag about their web-filtered-by-humans approach, as opposed to the "cold" algorithmic approach of Google. But it turns out Google is pretty human too, only with higher ambitions of creating generalizations from the human observations.
Wildcards in strings "apple * macintosh" will return pages with the word macintosh shortly following apple. Not reversable, but still quite useful for that kind of search.
Slashdot is as much of a blog as I am a Egyptian gerbil. Slashdot links to stories that generate discussions. Slashdot is NOT about the people that create the posts, but about the people that comment here.
My Starcraft 2 Blog
If the UK sites in particular are the ones you want out of you search results, compare these searches on Google:
digestives london
digestives london -inurl:.uk
Not sure about this:
"Google rarely allows outsiders to visit the unit, and it has been cautious about allowing Mr. Singhal to speak with the news media about the magical, mathematical brew inside the millions of black boxes that power its search engine."
I could see tens of thousands, maybe hundreds of thousands, but millions?
It is rather simple (I am an insider).
Google breaks pages in words. Then, for evey word it keeps a set which contains all the pages (by hash ID) that contain that word. A set is a data structure with O(1) lookup.
When you search for "linux+kernel" google just does the set union operation on the two sets.
Now a "word" is not just a word. In google sees that many people use the combination linux+kernel, a new word is created, the linux+kernel word and it has a set of all the pages that contain it. So when you search for linux+kernel+ppp we find the union of the linux+kernel set and the "ppp" set.
So every time you search, you make it better for google to create new words. And this is part of the power of this search engine. A new search engine will need some time to gather that empirical data.
Of course, there are ranks of sets. For example, for the word "ppp" there are, say, two sets. The pages of high rank that contain the word ppp, and the pages of low rank. When you search for ppp+chap, first you get the set union of the high rank sets of the two words, etc.
Now page rank has several criteria. Here are some:
well ranked site/domain, linked by well ranked page, document contains relevant words, search term is in the title or url, page rank not lowered by google emploee (level 1), page rank increased, etc.
It is not very difficult actually.
(posting AC for a reason).
So how do you call the "thing" that you use to impement a heuristic?
A classifier is a black box which takes some data as input, and computes one or more scores. The simplest example is a binary classifier, say for spam. You feed some data (eg an email) and you get a score back. If it's a big score say, then the classifier thinks it's spam, and if it's a small score it's not spam. More generally, a classifier could give three scores to represent spam, work, home, and you could pick the best score to get the best choice.
So you should really think of a classifier as a little program that does one thing really well, and only one thing. For example, you can build a small classifier that looks if the input text is english or russian. That's all it does.
Now imagine you have 100 engineers, and each engineer has a specialty, and each builds a really small classifier to do one thing well. The logic of each classifier is black boxed, so from the outside it's just a component, kind of like a lego brick. What happens when you feed the output of one lego brick to the input of another lego brick?
Say you have three classifiers: english spam recognizer, russian spam recognizer, english/russian identifier. You build a harness which uses the english/russian identifier first, and then depending on the output your program connects the english spam recognizer or the russian spam recognizer.
Now imagine a huge network with some classifiers in parallel and some classifiers in series. At the top there's the query words, and they travel through the network. One of the classifiers might trigger word completion (ie bio -> biography as in the article), another might toggle the "fresh" flag, or the "wikipedia" flag etc. In the end, your output is a complicated query string which goes looking for the web pages.
The key idea now is to tweak the choice thresholds. To do that, there's no theory. You have to have a set of standard queries with a list of the outputs the algorithm must show. Let's say you have 10,000 of these queries. You run each query through the machine, and you get a yes/no answer for each one, and you try to modify the weights so that you get a good number of correct queries.
Of course you want to speed things up as much as possible, you can use mathematical tricks to find the best weights, you don't need to go get the actual pages if your output is a query string you just compare the query string with the expected query string etc, but that would be depend on your classifiers, the scheme used to evaluate the test results, and how good your engineers are.
The point is that there's no magic ingredient, it's all ad-hoc. Edison tried a hundreds of different materials for the filament in his lightbulb. Google is doing the same thing according to the article. What matters for this kind of approach is a huge dataset (ie bigger than any competitors') and a large number of engineers (not just to build enough components, but to deprive its competitors of manpower). The exact details of the classifier components aren't too important if you have a comprehensive way of combining them.
And the thing that I want to know is how they evaluate the results. I actually do research in this space right now, and by far the most painful thing is evaluation of results. We have a system that automates most of the work, but there's still a lot of human involvement, and this limits the input dataset size and speed with which we can iterate the improvements.