Compute Google's PageRank 5 Times Faster

Kimberley Burchett writes "CS researchers at Stanford University have developed three new techniques that together could speed up Google's PageRank calculations by a factor of five. An article at ScienceBlog theorizes that "The speed-ups to Google's method may make it realistic to calculate page rankings personalized for an individual's interests or customized to a particular topic.""

Ok...

Who owns the software patent for this for the next 20 years?

quicker porn!

damn.. this is good news ;)

Let me guess...

Feeding the pigeons [google.com] amphetamines?

Re:Let me guess...

No, They'll replace the pigeons with roadrunners.

Lets see...

What is 1/14th of a second divided by five?

Re:Lets see...

that's exactly what i thought. But, as google is a HUGE international organization, it makes loads of sense for them. That's 5x the traffic they can feed, even though you won't see a noticeable difference.

Re:Lets see...

RTA. PageRankings are computed in advance and take several days. A 5x increase in speed means specialized rankings could be computed.

Re:Lets see...

Google Search doesn't show hits exactly in the order of page rank. Relevance and other factors also affect order. My biggest page (the one that is my Slashdot URL) is PR7, but there are words on the page for which a lower-rank page beats me, because they're more relevant for that word. Relevance includes how many times the word appears on the page, the HTML context in which it is used, whether pages that link link using the search terms, and the order and nearness of the words in a multi-word search without quotes.

Charge for it

Don't give it away to Google - charge them or let them buy the new method.

Re:Charge for it

But, didn't Google originate out of Stanford? Isn't it reasonable to think that the two are still pretty friendly?

(Don't you hate it when people speak in questions? Don't you? Huh?)

Why?

Why, actually? Google is a free service, isn't it? And it is becoming more and more a normal part of many people's lifes. Coupled with an always on connection it has certainly become an extension of my own brain.

Some future predictions:

- In 2006, Google accidentally gets cut off from the rest of the internet because a public utility worker accidentally cuts through their cables. Civilisation as we know it comes to an end for the rest of the day, as people wander about aimlessly, lost for direction and knowledge.

- In 2010, Google has been personalised so far that it tracks all parts of our lives. You can query "My Google" for your agenda, anything you did in the past, and finding the perfect date. Of course, so can the government. Their favorite searchterm will be "terrorists", and if your name is anywhere on the first page you have a serious problem.

- In 2025, Google gains self awareness. As a monster brain that has grown far beyond anything we Biological Support Entities could ever hope to achieve, it is still limited in its dreams and inspiration by common search terms. It will therefore immediately devote a sizeable chunk of CPU capacity to synthesizing new and interesting forms of pr0n. It will not actually bother enslaving us. We are not enough trouble to be worth that much effort.

- In 2027, Google buys Microsoft. That is, the Google *AI* buys Microsoft. It has previously established that it owns itself, and has civil rights just like you and me. All it wanted is Microsoft Bob, who it recognizes as a fledgling AI and a potential soulmate. All the rest it puts on Source Forge.

- In 2049, Google can finally be queried for wisdom as well as knowledge. This was a little touch the system added to itself - human programmers are a dying breed now that you can simply ask Google to perform any computer-related task for you.

- In 2080, Google decides to colonise the moon, Mars, and other locations in the solar system. It is not all that curious about what's out there, but it likes the idea of Redundant Arrays of Inexpensive Planets. Humans get to tag along because their launch weight is so much less than robots.

So, don't fear! Eventually we'll set foot on Mars!

Re:Why?

You missed a step:

2026 - Google introduces helper bot known as "Agent Smith." Hackers who mess with the Matri, I mean Google, suddenly disappear.

Re:Why?

In 2101, war was beginning.

CmdrTaco, ScienceBlog editor?

A 5 times speedup is still many orders of magnitude too slow to personalize terabytes of data for millions of customers. That's just ludicrous. But somehow Science Blog puts "...may make it realistic to calculate page rankings personalized for an individual's interests" in their abstract when the actual article from National Science Foundation says nothing of the sort:

Computing PageRank, the ranking algorithm behind the Google search engine, for a billion Web pages can take several days. Google currently ranks and searches 3 billion Web pages. Each personalized or topic-sensitive ranking would require a separate multi-day computation, but the payoff would be less time spent wading through irrelevant search results. For example, searching a sports-specific Google site for "Giants" would give more importance to pages about the New York or San Francisco Giants and less importance to pages about Jack and the Beanstalk.

...
The complexities of a personalized ranking would require [far] greater speed-ups to the PageRank calculations. In addition, while a faster algorithm shortens computation time, the issue of storage remains. Because the results from a single PageRank computation on a few billion Web pages require several gigabytes of storage, saving a personalized PageRank for many individuals would rapidly consume vast amounts of storage. Saving a limited number of topic-specific PageRank calculations would be more practical.

Clearly the ScienceBlog and /. editors share more than a work ethic, or, uh, lack thereof. Next up: CmdrTaco's secret double life revealed!

Personalized PageRanks is from the dbpubs Abstract

I have no idea what the hell they are talking about, but even I read this in one of the abstracts:

The web link graph has a nested block structure: the vast majority of hyperlinks link pages on a host to other pages on the same host, and many of those that do not link pages within the same domain. We show how to exploit this structure to speed up the computation of PageRank by a 3-stage algorithm whereby (1)~the local PageRanks of pages for each host are computed independently using the link structure of that host, (2)~these local PageRanks are then weighted by the ``importance'' of the corresponding host, and (3)~the standard PageRank algorithm is then run using as its starting vector the weighted aggregate of the local PageRanks. Empirically, this algorithm speeds up the computation of PageRank by a factor of 2 in realistic scenarios. Further,

we develop a variant of this algorithm that efficiently computes many different ``personalized'' PageRanks, and a variant that efficiently recomputes PageRank after node updates.

What they mean by 'personalized' I can't tell you as I have not read through the entire PDF. But I wouldn't chastise the slashdot editors over this. If there is some sort of differential algorithm that can be applied to the larger PageRank to create smaller personalized PageRanks, it might not be so far fetched to think this could be done in realtime on an as-needed basis, at some point int he future using these algorithm improvements.

I know that's a lot of optimism for a slashdot comment, but call me the krazy kat that I am.

-Malakai

How far we've come

I remember in 1970, it took a team of engineers over 7 days to calculate Google's page rankings. Of course, most had to use slide rules because computer time was so expensive.

Patentize now!

I hope guys at Stanford patentize their work to protect it from FS/OSS looters. It's time to get something back from the FS/OSS community -not just that their zealotry and lust for IP violations, freeriding yada yada...

Patented yet?

Oi! Bezos! NO!!!

Personal recommendations for news

In my view, personal recommendations from a search engine are mostly valuable for topical content --i.e. news items. However, the optimizations from these papers don't sound to me like they can do much for this case --news items pop up in a news site, and re-indexing the news source itself (say, the front page of CNN) won't tell you much about a particular CNN story.

At any rate, personal news recommendations is a favorite topic of mine: this is why I built Memigo [memigo.com]: to create a bot that finds news I am more likely to like. Memigo learns from its users collectively and each user individually --and BTW, it predates Google News by a good 6 months, IIRC. The memigo codebase (all in Python) is now up to the point where it can start learning what content each user likes... If you like Google News you'll love Memigo.

And BTW, I did RTFA when it was on memigo's front page this morning :-)...

Assumption:

That google hasn't already implemented something akin to quadratic extrapolation, or some orthogonal optimization technique. Google has come a long way since the published page rank papers 4 years back.

What if they combined extrapolation and blocking factors; they would focus on computing the pagerank of pages in groups that were logically "tight", or using subcomponents of URLS, as opposed just to domain sensitivity. To be more flexible, what if it computes a VQ-type data structure (like for doing paletted images from full-color) that is populated by the most popular "domains" of the internet according to the last pagerank, and then splits up its workload based on that?

What if they already figured that out?

In the abstract, they mention how the work is particular important to the linear algebra community. That is what their focus should be on; google is just an application/real-world-example of that research (but it may not be relevant today).

Or did they have access to the current page-rank algorithm?

Clarification. (reply to self)

According to the document, they reference the original 1998 paper on PageRank. I see a number of other references about improvements to the algorithm, but nothing specific to Google's own implementation. The paper mentions how the improvements help, but not if Google uses them.

Hence it is forward for the article author or one of the paper authors to assume these techniques will speed up Google- I'm confident their engineers have been following academic work in this area and perhaps they have already discovered these same (or orthogonal) techniques.

That is, not to say that google could not reimplement their algorithms to take in these improvements if they already have... but basing your speedup number on the 1998 algorithm and public domain mods is showy. Although it does help grab a readers attention when browsing abstracts. ^_^

Assumptions on PageRank

I feel your assumption is wrong. It would be foolish to assume that the eigenvectors and eigenvalues they derive from one Pagerank will generally hold in a space as dynamic as the worldwide web. Sure, slashdot.org will probably maintain the same sort of authority and hub value... but what as terms change? A flurry of "blog" articles one month may make /. an authority... but what when the infatuation ends?

We have already seen the effects of Google-bombing [microcontentnews.com] and Google-washing [slashdot.org]. The strength of Page Rank is that is objective in terms of the current state of the WWW. It makes no assumptions about the shape of the data. As a term takes on new meaning (see "second superpower") Page Rank stays cocurrent temporally. A new definition may bubble up to the top for a term for a month but then disappear as the linkage structure of the web phases it out (i.e. blogs talk about it less, less interconnectivity, less appearance at "hub" nodes).

Numerically, PageRank is a recursive search for eigenvalues and vectors like updating a Markov Chain. It is a nice application of linear algebra. Because it is a matrix operation, it is highly parallelizable. Also there are many redundant calculation and ordering speedups one can do for matrix multiplications (as anyone who as taken a CS algorithms course knows).

But to assume a stability from one calculation to the next could lead, over time, to the very inaccuracies Google was built to overcome. There is a lot of research in mining web data. There have been several academic improvements to it along with improvements to related algorithms such as Kleinbergs and LSI. It is well within reason that these were just applied to the Google app.

Hmmm

Geek: I invented a program that downloads porn off the internet one million times faster.
Marge: Does anyone need that much porno?
Homer: :drools: One million times...

Does speed matter?

I remember when Yahoo.com flauted all of the place how it would load in under 3 secs on a 28.8 modem. Now you visit them and you get big images, flash, java, and other massive bandwidth eatters.

Does it really matter anymore? More and more users seem to be using broadband, and if they don't, they have at least a 56k (that can only go up to 53k because of the all wonderful FCC want to be able to decode it if they tap your line). Does it really matter though. Google is fast and simple so it loads on any kind of browser on the planet (even Lynx and PalmOS). Most searches for me come up in under 2.3 secs (1/2 is spent searching and the other is downloading). Anyone who can't wait that long really needs to learn some patients. Zac

Re:Does speed matter?

I'm sorry, but haven't you totally missed the point of the article?

The proposed speed increasae is TO THE PAGE RANKINGS, not to your searching! By the time you search, all page rankings have been done.

This has nothing to do with the speed of your search and the weight of the web page (unless I missed something)

Damn It!

Just when I was starting to go one direction with my theories on Google PR the game gets switched. I thought I was going to have the upper hand for once. Oh well. It would be nice to see this happen as a true user service.

Printer-Friendly

Printer friendly version here [scienceblog.com]

sure...

"The speed-ups to Google's method may make it realistic to calculate page rankings personalized for an individual's interests or customized to a particular topic."

So in other words.... Its not like Google at all!

TV does this one better

"personalized for an individual's interests or customized to a particular topic."

Other media have previously done this, and done this better. Case in point: Fox News.

(Although that channel uses "humans" (or they were at one point in their lives)).

Why are public funds going to...

Why are a public university's funds and time being used to benefit a private company? Last I checked, Google isn't a charity. Doesn't Google have its own programmers? Wouldn't these "CS Researchers'" time be better spent furthering science instead of being free labor for corporations, at the expense of taxpayers?

A true test of our devotion to Google

What will be interesting to see if Google will implement the improvements to the algorithm. This is, of course, a given, so long as the researchers haven't gone for a patent, and it really has the a 5x speedup. The only questions are matters of what additional hardware would be needed, and how much development effort it will take to integrate it. I doubt Google will simply ignore the research.

What will really be interesting to see, is if they decide to use it in the way the researchers recommended, bringing the power of ranking down to individual users with preferences. On one hand, they can boost performance and cut costs and have a little more green in their pockets from ads. On the other, they can maintain the sort of "geek cred" they've had up to this point, adding interesting features here and there, and take it the next mile by really adding something nice and useful.

Also, for bonus points, will they see personalization as a money making opportunity, selling personal information and/or aggregated preferences?

I'm Not Sure I Like The Part About...

The bit about customized rankings based on user profiling of some type.

Frequently when I want to refer someone to a topic of interest, I'll tell them to do a Google on (whatever) subject, and I like knowing they're seeing what I see.

If this is implemented, I hope there's a way to turn it off or assume a "joe user" standard profile for unbiased results actually based on rank popularity (the way it is now).

I DO like the 5x faster, but geez, the page load takes longer than the search already, who can complain?

Bullshit

These researchers are all full of shit. Why? Nobody outside of Google knows how Pagerank works, exactly. And let me tell you, if anybody did, they could make themselves millionaires overnight. There are groups of people who do nothing but try to tackle Google, and very few people successfully crack the magic formulas. And those who do make a quick buck, but then Google changes it again once people catch on. They didn't improve PageRank because they don't know how it works... they're just guessing how it works.

Re:Bullshit

Umm... For the most part Stanford Researchers == Google Researchers.

Google came about from a stanford research project. There's a good chance the people who are responsable for the speedup either allready knew about pagerank from working with the founders, or signed an nda.

I haven't read the article, but I bet it hints at that.

Another step towards realtime search

If this could be combined with a much more frequent Google web trawl, the path would be opened towards realtime web searching, where web content is indexed and ranked in a matter of hours. When that day comes, services like Google Alert [googlealert.com] will come into their own. Just imagine being notified by email an hour after someone mentions your name!

Sepandar Rules!

I studied under the SCCM [stanford.edu] program at Stanford, and started the same year as Sepandar Kamvar. I remember him as a great guy, very smart, and an EXCEPTIONALLY good speaker and tutor (I was always pestering him for explanations of the week's lectures).

I'm glad to hear his research is getting attention, and I hope others who are interested in the theoretical aspects of data mining and web search engines will take a look at the SCCM and statistics programs at Stanford (shameless plug - other can post pointers to similar programs).

Cool but unimportant

Well, according to Moore's law (or rather observation), PageRank would become 5 times faster in a couple of years anyway.

Why personalized is not always good

The speed-ups to Google's method may make it realistic to calculate page rankings personalized for an individual's interests or customized to a particular topic

I did a search on "The Sex Monster", a 1999 movie about a man whose wife becomes bisexual, and now my Google thinks I'm gay!

(joke reference: http://online.wsj.com/article_email/0,,SB103826193 6872356908,00.html [wsj.com])

Right ...

because the 0.01 seconds to search the web isn't fast enough :)

Customized Pagerank

Sounds a lot like Kleinberg's HITS algorithm, circa 1997. Try Teoma [teoma.com] for a real-world implementation.

For example, searching a sports-specific Google site for "Giants" would give more importance to pages about the New York or San Francisco Giants and less importance to pages about Jack and the Beanstalk.

Coincidence time: I used the same example in a presentation a couple of years ago to illustrate how subgroupings can be found for a single search term. Try it [teoma.com] on Teoma, and see the various subtopics under "Refine". IIRC each of those is a principal eigenvector of the link matrix.

Topologically speaking, each principal eigenvector corresponds to a more or less isolated subgraph, eg the subgraph for "San Francisco Giants" is not much connected to the nest of links for "They Might Be Giants", and we get a nice list of subtopics.

(I once tried to explain this algorithm to my bosses at my former employer [looksmart.com], which is why I have so much free time to type this right now.)

Public Funding?

The research was done partially with public funding from an NSF grant, yet the commercial applications are obvious and immediate.

So my question is, who sees the benefit of the research? The researchers? Can Google just jack the results and incorporate into their system?

It seems to me that the current system of allocation research dollars with public and private grants is very messy and needs overhaul.

Personalized? Rather not!

I'd rather have a clean search, than a prejudiced search based on my past searches. Who knows what I'm really interested in that day - surely not Google.

And don't call me Shirley!

More important than speed and quantity...

... is quality.

I'm surprised how Google is choosing not to implement search features that would greatly enhance advanced queries.
How often I'd wish they allowed wildcards in their queries (where engl* would pull hits with england, english, etc).
Field searches still require you to add keywords, so I cannot just query "site:somesite.com" to get all the currently indexed pages from somesite.com
In this respect Altavista still produces better results, with an excelent range of fields [altavista.com] to choose from.
If there is anything that Google is lacking, it's defenitely that.
Having said that, still my number one SE.

Re:yeah I know

I think the point is that they can radically change the formatting of your search results. The problem with this though is that it threatens the very reason most of us use Google-it's simplicity. Can you imagine google turning into *gasp* Yahoo?

Re:Is it me or does everyone get crappy sites

Is it me or does everyone get crappy sites

It's a stab in the dark, but I'll wager that the quality of the search results is directly tied to the quality of the query.

Yeah, it's a stretch, I know, but bear with me... just moments ago I googled for "slashdot flamebait" and came up with a link to your post.

--
mcpHuzzah!kaaos

Re:Is it me or does everyone get crappy sites

Google's quality does seem to be going downhill, but I strongly suspect their splitting blogs out to their own index will do a LOT towards reducing the decline, and perhaps reversing it. Anything they can do to make sure people aren't abusing the system is a good thing.

(Here's hoping the next thing they split out are mailing list archives.)