chung wrote:
> On Fri, 2010-06-18 at 15:25 +0300, Nina Jeliazkova wrote:
>   
>> chung wrote: 
>>     
>>> On Fri, 2010-06-18 at 15:07 +0300, Nina Jeliazkova wrote:
>>>   
>>>       
>>>>> Is there any kind of information in the representation of a model which
>>>>> should be hidden/protected? I think the model representations as they
>>>>> are now do not have anything about the model itself such as its
>>>>> equation, coefficients etc...
>>>>>   
>>>>>       
>>>>>           
>>>> No, but you have this in PMML / TEXT representation .  I would say the
>>>> current model API is a bit inconsistent, since RDF is actually a
>>>> metadata of the model, not the model itself.  
>>>> It might be reasonable to introduce additional URIs to retrieve model
>>>> content, where applicable, e.g.  /model/1 retrieves metadata,
>>>> but /model/1/content retrieves the model itself. 
>>>>     
>>>>         
>>> If you need model's metadata, ask for RDF otherwise ask for
>>> application/xml and you get the PMML representation of the model, i.e.
>>> its content. So I don't think we really need to separate these. What is
>>> more there are models with no content such as leverages or even some
>>> lazy ones (like kNN), so /model/{id}/content will be redundant.
>>>   
>>>       
>> This is true in case policies are not involved - but policies don't
>> account for media type - there is no way to protect PMML , but allow
>> RDF representation.
>>     
>
> Then just deny access to both RDF and PMML for a specific model. There
> is no point in allowing the user to see which independent features are
> involved in the model but not provide its PMML representation. Isn' it?
>   
It might make sense in some cases,  but more interesting question is -
what about the policies for the features?  Who/when and what policy is
assigned to them?

> If it is, then we have to separate model meta from model content. But
> again what about models with no content like some DoA?
>   

All models have some content, for lazy learning models it might be the
entire dataset.   If you mean applicability domain, based on leverage,
then it's the hat matrix and the threshold.

Regards,
Nina


> Best regards,
> Pantelis
>
>   
>> Regards,
>> Nina
>>     
>>> Best Regards,
>>> Pantelis
>>>   
>>>       
>
>