There is now a plugin interface in libferris so that the filesystem can use AI to make suggestions as to what emblems a file should have. Currently I have Bayesian (bogofilter) and Support Vector Machine (svm_light) agents. The interface supports anything that can train on selected files and the attachment of emblems and then proffer a fuzzy assertion or retraction given an example file. The agent’s beliefs are all collected and resolved to give an overall fuzzy assertion/retraction belief based on how trusted you think the agents are (the trust resolution should allow many resolution paths in the next release).
more details anywhere? Seems to be a bit light on that ferris link about this specifically..
Here:
http://www.osnews.com/story.php?news_id=1304
The API is in Agent.hh and plugins use Agent_private.hh.
the fagent command line tool is used to train and use agents.
<p>
There is a little snippit describing the screenshot for agents
http://witme.sourceforge.net/libferris.web/research/shots.html .
<p>
I’ve run the SVM agent through 200 positive and 200 negative example newfeeds from http://www.daviddlewis.com/resources/testcollections/reuters21578/ reuters-21578
and even with that small level of learning (the newfeeds are rather small text files) it got 5/6 right with the 1 it got wrong having an assertion level so you could tell that it was not confident that it was correct.
<p>
Patches will be accepted from folks who wrap other text classifiers or agents. The two current implemenations offer a decent starting point for folks wanting to add more. The SVM plugin handles creating RDFreq FeatureSets, caching FeatureSets, scaling by IDF vectors and normailzing FeatureSets.
<p>
If anyone knows of good code that exploits interword dependancy please post links