Martin Guetlein wrote:
> Hello All,
>
> Tobias Girschick wrote:
>   
>> Hello Nina,
>>
>> On Mon, 2009-12-21 at 14:21 +0200, Nina Jeliazkova wrote:
>> [...]
>>
>>     
>>>> RDF representations, structurally, contain much more (meta)information
>>>> about the objects they describe than ARFFs, so this piece of
>>>> information in the text/x-arff (the datatype of each feature) IMHO has
>>>> to be included in the RDF or at least - in order not to modify the RDF
>>>> standards we adopted in API 1.1 - we should use proper XSD datatypes
>>>> for every value. After all, its not 1^^double, 1^^string and
>>>> 1^^nominal is not the same and won't (shouldn't) be handled the same
>>>> way by a training algorithm.
>>>>
>>>>         
>>> Yes, especially for nominals, it would be better to introduce subclass
>>> of Feature, rather than using XSD types for denoting the types.  I might
>>> try to extend opentox.owl next days.
>>>
>>>       
>> This would be great. At the moment, classification is not possible as it
>> relies on a nominal target feature. Will this be reflected in the
>> text/x-arff, too?
>>     
>
> Apart from feature value type I would like to have a "feature range"
> as well. This should be a property of a feature, which gives me all
> the possible feature values of a nominal feature (e.g. active,
> moderately-active, inactive). This is needed when validating a
> prediction algorithm.
>
> What do you think?
>   

Yes, we might extend  opentox.owl class with a NominalFeature , a
subclass of Feature and introduce a property like ot:nominalValues ,
which allows specifying set of String values. If a dataset includes such
NominalFeature , one would be able to retrieve set of values ,that are
allowed by inspecting values of the ot:nominalValues property. 

This will be more or less equivalent to Weka nominal attributes, where
the allowed values are listed in the ARFF file header .

Best regards,
Nina
> Best regards,
> Martin
>
>
>