I've recently released a Ruby port "Bishop" of the "Reverend" bayesian
classifier written in Python. Bishop-0.3.0 is available as a Gem and
from RubyForge

http://rubyforge.org/projects/bishop/

Bishop is a reasonably direct port of the original Python code, bug
reports and suggestions for improving the structure of the code would
be welcomed.

Bishop includes both Robinson and Robinson-Fisher algorithms for
classification.  It is presumed that they were correctly implemented
in Reverend.  I aim to test this in my own use of the code.

Support is included for saving/loading the trained classifier to/from YAML.

An example of using Bishop:

  require 'bishop'

  b = Bishop::Bayes.new
  b.train( "ham", "a great message from a close friend" )
  b.train( "spam", "buy viagra here" )
  puts b.guess( "would a friend send you a viagra advert?" )

  => [ [ "ham", <prob> ], [ "spam", <prob> ] ]

Bishop defaults to using the Robinson algorithm.  To use a different
algorithm construct the classifier passing a block which will call the
choosen algorithm:

  Bishop::Bayes.new { |probs,ignore| Bishop::robinson_fisher( probs, ignore ) }

To save to a YAML file:

  b.save "myclassifier.yaml"

To load from a YAML file:

  b.load "myclassifier.yaml"

You can uniquely identify training items

  b.train( "ham", "friends don't let friends develop on shared
hosting", "<xyz.323-ON8.002-802....@ham.com>" )

An can untrain items:

  b.untrain( <pool>, <item>[, <uid> ] )

I'm using this in a project of my own and would welcome any feedback
or suggested improvements.

Regards,

Matt

#I've recently released a Ruby port "Bishop" of the "Reverend"
#bayesian classifier written in Python. Bishop-0.3.0 is
#available as a Gem and from RubyForge
#
# http://rubyforge.org/projects/bishop/

hmmm, another cool filter. very small, took me less than 5 seconds to
install remotely the gem.

btw, matt, how difficult or easy it it to port the bishop database to a db
like postgres? I am asking since i may be querying/archiving more than
10_000 entries...

On 4/15/05, "Peña, Botp" <b...@delmonte-phil.com> wrote:

The problem I have is that the classifier will need to be available to
process each request classifying an item and each request that sorts
items, i.e. quite often.  This probably means initializing it once and
storing it in a session variable.  Since there is no concept of a
session expiry callback (Rails is not an app server).  The question is
"How do I checkpoint the classifier as it is trained?"

At the moment I can serialize it to YAML but that takes a little time
and will get slower as the training set increases.  Doing the YAML
conversion and a SQL update on each request is prohibitive.

I've been considering whether to have the classifier exist in a
separate thread|process and allow it to checkpoint itself
automatically at intervals independent of the users session behaviour.
 Another option was to convert the code so that everything (or nearly
everything) operated directly out of the database.

Representing the pools and training data via SQL would be simple
enough since it's just (word,count) tuples.  Basing the code on a SQL
variant might be quite attractive.  The issue would be making it
portable.

Since I'm using Rails anyway I could certainly attempt an ActiveRecord
based variant which should satisy the Postgres requirement also.

Could be an interesting experiment.  What do you think?

M

interesting indeed, and useful too.
thanks and kind regards -botp

It looks like they both have a similar syntax:

classifier.train :symbol, "content"

Would the method_missing syntax be easy to add to Bishop? Would
untrain be easy to add to projects/classifier? From what I've looked
at them so far, sounds like the answer to both would be yes. If they
had the same API, they could go in the same module so that swapping
filter types would be as simple as changing the Classifier::XXX.new
line.

Thank you for this useful librbary.
I am trying to use it to analyse the project of text for the european
constitution (Is it social? liberal? respectful of human rights?) I am
doing this for myself, just out of curiosity, there is no responsibility
or any liability involved in the usage of the classifier or in the result.
I'd like to know what the behaviour of the training of a classifier is
when two different set of words are submitted in two successive "train"
method invocations for a given category. Does the second invocation
resets the training or does it accumulate the "experience" progressively.

If I had known Lucas was working on his classifier library before I
did the port of Reverend I probably wouldn't have bothered.  However I
have done it and am using it in another project of my own and have had
some ideas about possible future developments.

One example is to build a version which runs directly from a SQL
database (possibly using ActiveRecord).  I'm also interested in new
algorithms and possible improvements for support classifying RSS items
within a tag space.

None of which precludes rolling Bishop and Classifier into one project.

However right now I'd like to keep control of Bishop and not be
constrained from making possibly incompatible changes to the API or
implementation.  Similarly Lucas may have his own plans for how he
wants to see Classifier develop.

I don't see the harm in having two projects and what I've suggested to
Lucas is that we should compare notes periodically and see if it makes
sense to merge the projects.  I guess also if a lot of users of the
libraries made a fuss this would affect my opinon.

Regards,

Matt

On 4/18/05, Jaypee <rf.ooda...@sd.eepyaj> wrote:

There is an #untrain method if you want to remove previously trained
information.

Regards,

Matt