Hi Martin,
By incorporating PageRank into the decision of what pages to crawl,
CommonCrawl is actually trying to approximate what search engine
crawlers are doing. In general, search engines would collect pages that
would be more likely to rank higher in search results, and PageRank is
an important component of that.
PageRank does transfer along the edges of the web graph, so a highly
ranked homepage would transfer it's PageRank to the pages leading from it.
My only complaints about CommonCrawl in this regard is that they don't
publish their webgraph and the computed scores. It's a valuable resource
to have. Further, they should compute it regularly... it seems they have
two dumps with two years apart, and if they used the PageRank scores
from the first dump to crawl the second, that might be a bit off.
Cheers,
Peter
On 4/17/12 3:25 PM, Martin Hepp wrote:
Dear Chris, all,
while reading the paper [1] I think I found a possible explanation why
WebDataCommons.org does not fulfill the high expectations regarding the
completeness and coverage.
It seems that CommonCrawl filters pages by Pagerank in order to determine the
feasible subset of URIs for the crawl. While this may be okay for a generic Web
crawl, for linguistics purposes, or for training machine-learning components,
it is a dead end if you want to extract structured data, since the interesting
markup typically resides in the *deep links* of dynamic Web applications, e.g.
the product item pages in shops, the individual event pages in ticket systems,
etc.
Those pages often have a very low Pagerank, even when they are part of very
prestigious Web sites with a high Pagerank for the main landing page.
Example:
1. Main page: http://www.wayfair.com/
--> Pagerank 5 of 10
2. Category page: http://www.wayfair.com/Lighting-C77859.html
--> Pagerank 3 of 10
3. Item page:
http://www.wayfair.com/Golden-Lighting-Cerchi-Flush-Mount-in-Chrome-1030-FM-CH-GNL1849.html
--> Pagerank of 0 / 10
Now, the RDFa on this site is in the 2 Million item pages only. Filtering out
the deep link in the original crawl means you are removing the HTML that
contains the actual data.
In your paper [1], you kind of downplay that limitation by saying that this approach yielded "snapshots
of the popular part of the web.". I think "popular" is very misleading in here because the
Pagerank does not work very well for the "deep" Web, because those pages are typically lacking
external links almost completely, and due to their huge number per site, they earn only a minimal Pagerank
from their main site, which provides the link or links.
So, once again, I think your approach is NOT suitable for yielding a corpus of
usable data at Web scale, and the statistics you derive are likely very much
skewed, because you look only at landing pages and popular overview pages of
sites, while the real data is in HTML pages not contained in the basic crawl.
Please interprete your findings in the light of these limitations. I am saying this so
strongly because I already saw many tweets cherishing the paper as "now we have the
definitive statistics on structured data on the Web".
Best wishes
Martin
Note: For estimating the Pagerank in this example, I used the online-service
[2], which may provide only an approximation.
[1] http://events.linkeddata.org/ldow2012/papers/ldow2012-inv-paper-2.pdf
[2] http://www.prchecker.info/check_page_rank.php
--------------------------------------------------------
martin hepp
e-business& web science research group
universitaet der bundeswehr muenchen
e-mail: h...@ebusiness-unibw.org
phone: +49-(0)89-6004-4217
fax: +49-(0)89-6004-4620
www: http://www.unibw.de/ebusiness/ (group)
http://www.heppnetz.de/ (personal)
skype: mfhepp
twitter: mfhepp
Check out GoodRelations for E-Commerce on the Web of Linked Data!
=================================================================
* Project Main Page: http://purl.org/goodrelations/