commons-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Phil Steitz <phil.ste...@gmail.com>
Subject Re: (MATH-607) Current Multiple Regression Object
Date Wed, 06 Jul 2011 18:03:55 GMT
On 7/6/11 10:37 AM, Greg Sterijevski wrote:
> I like the following:
> It is conceivable that the interface you designed could be a facade over a
> lower level linear operator interface where that makes sense.  If so, that
> is great.
>
> Looking through commons there is public interface DecompositionSolver.
> Perhaps an extension of this interface is what you are thinking of?
>
> Phil what are your thoughts?

Sorry to be slow to respond on this.  I am not sure I understand
exactly what Ted is getting at, as it would seem that at least from
the model definition / data acquisition standpoint, what you have
defined is about as low-level as you can get and appropriate for the
use case, which is to support incremental adding / streaming of data
into a multiple regression model (like the "storeless" statistics
elsewhere in  [math]).  Enabling large models to be specified via
linear operators as well would be good if that is the suggestion,
but it is not clear to me how a linear operator model specification
interface could support the use case that the MATH-607 is addressing.

I have some more specific comments on the patch that I will add to
the ticket.

Thanks!

Phil
> -Greg
>
> On Wed, Jul 6, 2011 at 11:54 AM, Ted Dunning <ted.dunning@gmail.com> wrote:
>
>> It isn't really necessary to commingle approaches at all.  It is just nice
>> to think about the alternatives at once to get better designs.
>>
>> It is conceivable that the interface you designed could be a facade over a
>> lower level linear operator interface where that makes sense.  If so, that
>> is great.
>>
>> If not, the question that is interesting is whether it could be slightly
>> adjusted to fit that style.
>>
>> On Wed, Jul 6, 2011 at 9:46 AM, Greg Sterijevski <gsterijevski@gmail.com
>>> wrote:
>>> I see. Why would it be a good reason to commingle functionality? Aside
>> from
>>> diagnostics like condition numbers and maybe eigenvalues, these
>> approaches
>>> don't seem to share much commonality. I could be wrong since my knowledge
>>> of
>>> Mahout style problems is a bit spotty.
>>>
>>> On Wed, Jul 6, 2011 at 11:34 AM, Ted Dunning <ted.dunning@gmail.com>
>>> wrote:
>>>
>>>> The other way that regression is done at scale is with a linear
>> operator.
>>>>  This linear operator is often defined by the behavior of some external
>>>> system that is not susceptible to incremental construction.  A good
>>> example
>>>> is a large text retrieval system.
>>>>
>>>> It would be useful to support that style interface as well.
>>>>
>>>> On Wed, Jul 6, 2011 at 9:29 AM, Greg Sterijevski <
>> gsterijevski@gmail.com
>>>>> wrote:
>>>>> Borrowing liberally from the SimpleRegressionClass,  the above
>>>>> functionality
>>>>> describes most of what a user would expect from a classical
>> regression
>>>>> analysis. What the interface buys us is the ability to support the
>> many
>>>>> ways
>>>>> to generate the results above: QR factorizations, in place gaussian
>>>>> elimination, incremental SVD and so forth.
>>>>>


---------------------------------------------------------------------
To unsubscribe, e-mail: dev-unsubscribe@commons.apache.org
For additional commands, e-mail: dev-help@commons.apache.org


Mime
View raw message