Wiki Mines

NickKnowles and I spend a few days looking at how wiki pages are linked. We wrote a small object model for pages and the links between them. This was coded in SqueakSmalltalk and populated from a file we first extracted from the wiki database. Once in memory we could write ad hoc queries against this model and expect them to be answered in seconds. Here are a few of our initial observations ...

I know none of this comes as a surprise. For that reason I invite you to try your hand at mining wiki to see what you can discover. Here is a zip of the text file we used to load our application ...

When unzipped this produces a single file, links.txt, that contains one line for each wiki page. The first word of the line is the page name. Subsequent words are links leading from the page. Additional punctuation indicate the presence of horizontal rules and signature indications. Here is a sample line for a particularly short page ...

  ColorOutsideTheLines --RaySchneider | KnowTheRule --RonJeffries
Please have a look at this data and tell us what you find. -- WardCunningham

The above zip file has not been updated since June 1999. For current links info, use http://c2.com/wiki/links.txt, which is updated nightly.
Jeff Grigg's quick analysis of the links text from above:
  559 of 4970 pages (11%) have "Xp" or "Extreme" in title or content references.
(This includes a few "false positives" like ActiveXpert? and XpFreeZone, but looks pretty close to being "right.") On second thought, I'm pretty sure it underestimates XP influence, as there are a number of pages, like "DoWhatYouSaySayWhatYouDo" which contain references to XP concepts (CodeUnitTestFirst in this case), but do not contain "Extreme" or "Xp". Tracking these down would be tedious.
  498 pages with "Extreme," 158 with "Xp" (= 17% overlap).
Analysis done with Vim, a more featureful "vi" clone with a Windows version.

I found 760 lines containing the word "patterns" out of the 4970, so I nominate that patterns still has more pages than XP. (Analysis done with MS Word search and replace facility and line count tool.) -- AlistairCockburn
I'll have to look for the cite, but there was an article in CACM a few months ago about using link structure to mine the web. -- MichaelFeathers
See also WikiReadingHabits for more insight on Wiki usage.

The technology behind Netscape's "what's related" button ...

Alexa maintains a multi-terabyte central database that stores both the behavioral patterns of Web travelers and categorical information about the actual content on the Web, the latter of which surfaces Site Statistics about any page on the Web. Related Links are generated using sophisticated data mining techniques along with intelligent technologies, both of which identify usage patterns and the relationships between the pages, based on common hypertext links and similarities in textual content.
The technology in alexa & the bots is ingenious and powerful, but mining the Web in the large is a different (and harder) problem than mining a wiki.

Wiki ought to have certain advantages when it comes to analysing the links and the link traffic for semantic insights. For example My impression is that it would be quite possible to have indexing bots operating over a wiki that would be able to cluster the the pages into useful RoadMap like indexes pretty well automatically. -- NickKnowles

True, but some of the ideas might still be useful in this domain, or at least suggestive of other strategies. Here's a nice article from last month's Scientific American that contains some very good ideas: http://www.sciam.com/1999/0699issue/0699raghavan.html. -- GlennVanderburg


Further reading:

A wiki seems like a very appropriate thing to run through a WebSom (see http://websom.hut.fi/websom/) because all the pages are in a fairly narrow universe of discourse, mostly in the same language, and consist almost entirely of relevant text. The websom doesn't take into consideration the link structure, though, and something tells me that's crucial to capturing information about a wiki. Also, some of the WikiClones that store (or can recover) meta-information, like who edited what and when, could look for "dialog" patterns, ie. Tom always responds to Mary's comments. -- SeanScoggins?


	Wiki stats for May 27, 2000

Num pages: 9622 Size sum: 19,904,634 bytes Mean size: 2069 bytes

3986 different page sizes. Min size: 1 byte (6 pages this size) Max size: 150827 bytes (1 page) Median size: 767

Size distribution % size 1 16 2 21 3 25 4 29 5 34 10 67 15 111 20 166 25 231 30 310 35 392 40 502 45 618 50 767 55 965 60 1190 65 1494 70 1854 75 2324 80 2989 85 3917 90 5295 95 7878 96 8846 97 10720 98 13385 99 17842

16 largest pages: # size 9607 30000 9608 30038 9609 30316 9610 30350 9611 31059 9612 32288 9613 33565 9614 35564 9615 36237 9616 36353 9617 36780 9618 46070 9619 88372 9620 106063 9621 122727 9622 150827

Ahh, yes! Now I get it. It's all about digging Wiki <ahem>, as opposed to planting explosives here. Okay.
See: WikiWordStatistics

CategoryWiki, CategoryWikiStructure

EditText of this page (last edited March 22, 2006) or FindPage with title or text search