[ANN] Bishop 0.3.0 - bayesian classifier for Ruby ported from Python

[Matt Mower]

Hi folks,

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", "<[email protected]>" )

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

 

[Douglas Livingstone]

Hi folks,

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

Could this be combined with http://rubyforge.org/projects/classifier/ ?

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.

Cheers,
Douglas

 

[gabriele renzi]

Douglas Livingstone ha scritto:

Could this be combined with http://rubyforge.org/projects/classifier/ ?

+1 on this question/suggestion.
There may be reasons to have two different libraries, but IMVHO it
would be better to have one slightly bigger library sharing APIs,
services and keeping the useful differences.

 

[Jaypee]

Matt Mower a écrit :

Hi folks,

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 ....

Regards,

Matt

Hello Matt,

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.

Thanks again ...
Jean-Pierre

 

[Matt Mower]

Douglas Livingstone ha scritto:

+1 on this question/suggestion.
There may be reasons to have two different libraries, but IMVHO it
would be better to have one slightly bigger library sharing APIs,
services and keeping the useful differences.

I thought it was about time I responded to this.

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

 

[Matt Mower]

Hi Jean-Pierre,

Thank you for this useful librbary.

You're welcome.

I am trying to use it to analyse the project of text for the european
constitution (Is it social? liberal? respectful of human rights?)
[..snip..]
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.

You're right when you say it accumulates. Further training supplies
more evidence to the classifier about which words are associated with
which categories . It uses this evidence to work out conditional
probabilities which are then combined to make a guess about the
approriate category for an item.

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

Regards,

Matt

 

[gabriele renzi]

Matt Mower ha scritto:

<snip all>
thanks for taking time to answer, I can understand your reasons and I'm
glad to know there is at least a touch beetween different hackers on
similar projects, thanks both

 

[Jaypee]

Matt Mower a écrit :

Hi Jean-Pierre,

Thank you for this useful librbary.

You're welcome.

I am trying to use it to analyse the project of text for the european
constitution (Is it social? liberal? respectful of human rights?)
[..snip..]
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.

You're right when you say it accumulates. Further training supplies
more evidence to the classifier about which words are associated with
which categories . It uses this evidence to work out conditional
probabilities which are then combined to make a guess about the
approriate category for an item.

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

Regards,

Matt

Thank you,
Jean-Pierre

 

[Lucas Carlson]

The subversion trunk of projects/classifier (see
http://rufy.com/svn/classifier/trunk) has the untrain method in it.
This will be released soon under Classifier 1.2.