Curiouser and Curiouser! on memescope http://matt.blogs.it/ RSS feed for topic memescope Copyright 2006 Matt Mower Squib/0.1 self@mattmower.com self@mattmower.com en-gb Bayesian sillyness on a Sunday afternoon Sun, 24 Apr 2005 22:25:29 +0100 <p><a href="http://www.memescope.com/">MemeScope</a> is a project i've been working on for a little while now. The aim is to research ideas about RSS, tagging, and information management within a personal context. Working in <a href="http://www.ruby-lang.org/">Ruby</a> and utilizing the <a href="http://www.rubyonrails.com/">Rails</a> framework allowed me to make quick progress on the basic application however, lately, I've found it hard to make time with work and study (exams in May -- groan) pressing in on me. Today though I made time to do a little experiment using <a href="http://rubyforge.org/projects/bishop/">Bishop</a>. This has been done very quickly and crudely but I hope it will at least indicate a useful path to follow.</p> <p>Searching my bloglines subscription page I grabbed the first bunch of RSS2.0 feeds I could lay my hands on. They were (in the order in which I could remember whether these people used vanilla RSS2.0): <a href="http://paolo.evectors.it/">Paolo</a>, <a href="http://matt.blogs.it/">me</a>, <a href="http://radio.weblogs.com/0001011/">Scoble</a>, <a href="http://www.scripting.com/">Dave Winer</a>, <a href="http://www.hyperorg.com/blogger/">David Weinberger</a>, <a href="http://marc.blogs.it/">Marc Canter</a>, <a href="http://www.jackvinson.com/">Jack Vinson</a>, <a href="http://www.terryfrazier.com/weblog/">Terry</a>, and <a href="http://blog.mathemagenic.com/">Lilia</a>.</p> <p>I created a classifier and trained it with the feed of each person, using their name as a pool. Then I fed another couple of blogs through and looked at the guessing results.</p> <p><a href="http://www.problogger.net/">Problogger</a>:<ul> <li>Jack = 0.468318829421744</li><li>Lilia = 0.00342010541710014</li><li>Marc = 0.282793689816727</li><li>Me = 0.393100714237438</li><li>Paolo = 0.346225200485668</li><li>Scoble = 0.253258334183292</li><li>Terry = 0.22771214945352</li><li>Weinberger = 0.0167048658202115</li><li>Winer = 0.289825786625603</li></ul></p> <p><a href="http://radio.weblogs.com/0110772/">Seb Paquet</a>:<ul><li>Jack = 0.576259503998576</li><li>Lilia = 0.235407686747673</li><li>Marc = 0.417682412318975</li><li>Me = 0.30854512908974</li><li>Paolo = 0.397301832715516</li><li>Scoble = 0.324509204588817</li><li>Terry = 0.291899188854501</li><li>Weinberger = 0.283141685408383</li><li>Winer = 0.108271483644474</li></ul></p> <p><a href="http://seblogging.cognitivearchitects.com/">Sebastian Fiedler</a>:<ul><li>Jack = 0.5</li><li>Lilia = 0.24819059477181</li><li>Marc = 0.5</li><li>Me = 0.234146370920656</li><li>Paolo = 0.326586572303767</li><li>Scoble = 0.317525809666623</li><li>Terry = 0.344058594866734</li><li>Weinberger = 0.208336788611639</li><li>Winer = 0.203113831786535</li></ul></p> <p>The numbers here are probabilities indicating (as far as the classifier is concerned) the likely hood of the guessed feeds being the same as the sample feeds. There is probably far too little data here to determine anything useful however, given the people concerned, I do find something suggestive about the results. Enough to make me want to look harder.</p> <p>For calibration here are two other results. Feeding Paolo's feed back in for guessing:<ul> <li>Jack = 0.162912565311238</li><li>Lilia = 0.00398573502931349</li><li>Marc = 0.0498689666595031</li><li>Me = 0.00988167948078622</li><li>Paolo = 0.5</li><li>Scoble = 0.0404166297297532</li><li>Terry = 0.0308895504863602</li><li>Weinberger = 0.0169976173877616</li><li>Winer = 0.0083263139102045</li></ul> and for my own feed:<ul><li>Jack = 0.355604304977086</li><li>Lilia = 0.00691730756404335</li><li>Marc = 0.148014510140305</li> <li>Me = 0.5</li><li>Paolo = 0.0873206208996655</li><li>Scoble = 0.0863911457566834</li><li>Terry = 0.105989792649879</li><li>Weinberger = 0.0781196567440771</li><li>Winer = 0.00582495354516671</li></ul></p> <p>Next time round I'm going to be more thorough, allow the use of RSS1.0 and Atom feeds, think about adjustments for text quantity and experiment with the use of the Robinson-Fischer algorithm for combining probabilities.</p> http://matt.blogs.it/entries/00002529.html