Re: Numerical attributes in NaiveBayes

On Tuesday, September 23, 2003, at 09:12  AM, Dominique Vlieghe wrote:

> On Tue, 2003-09-23 at 16:00, Ken Williams wrote:
>>
>> Algorithm::NaiveBayes uses numerical attributes:
>>
>>    $nb->add_instance
>>      (attributes => {foo => 1.7, bar => 3.234},
>>       label => 'whatever');
>>
>> Or do I misunderstand your question?
>
> I think you do, what I mean is that it should create some sort of
> distribution (e.g. a normal distribution) of the values of a given
> attribute. The naive bayes modules will (if I'm not mistaken) count the
> number of occurrences of a (numeric or not) attribute e.g. in your
> example you will have 1 times 1.7 and 1 times 3.234.

Right, those values essentially mean that the "foo" attribute is 
counted 1.7 times, and the "bar" attribute is counted 3.234 times.  
They're simply arbitrary weights, so "count" is not quite the right 
term, but it's the way many people think of them.

So what you're looking for is something like this, right?

    $nb->attribute_type('numeric');
    $nb->add_instance
      (attributes => {23 => 1.7, 35 => 3.234},
       label => 'whatever');

and then the model should correctly interpolate values for unseen 
attributes like 29 using an underlying distribution model?  Currently 
this is unsupported by the module, but it could be added - especially 
with some help. =)

  -Ken