The example in the FAQ is a simple case where you have only one time-tube
(say, one store with one title). However, the idea can be generalized to
support N-dimensional time-tube with nested table concept. The core concept
is that (1) each individual time-point (time-slice) becomes a case, (2)
introduce Minus-i columns up to how much previous time-points you want the
model to look at to predict now, (3) use nested table key to model multiple
dimension. Here is an example:
CREATE MINING MODEL myTSModel
(
Year LONG KEY,
SalesTrends TABLE
(
Store TEXT KEY,
Title TEXT KEY,
SaleNow DOUBLE DISCRETIZED(50) PREDICT,
SaleMinus1 DOUBLE,
SaleMinus2 DOUBLE,
SaleMinus3 DOUBLE
)
) USING Microsoft_Decision_Trees
This model will be able to predict SaleNow of any combination of Store/Title
based on any combination of Store/Title and their 3 previous Sale amounts.
Note that you should turn-off automatic feature selection by setting
MAXIMUM_INUT_ATTRIBUTES and MAXIMUM_OUTPUT_ATTRIBUTES to a huge integer.
However, being an indirect support (or I should say a trick), this model has
several potentially serious problems:
- SQL Server 2000 Microsoft_Decision_Trees doesn't support continuous
attribute being predicted. As a workaround, we used DISCRETIZED in this
example. This may adversely affect the prediction accuracy.
- This model won't be able to learn any periodicity in prediction. The
algorithm is not time-aware.
- This trick introduces huge number of attributes, which would have been
much simpler if the algorithm was aware of time-tube concept. For instance,
automatic feature selection capability in Microsoft_Decision_Trees is not
helpful since it's not time-aware.
Again, in Yukon beta, all the issues have been taken care of and the model
definition has become intuitive and simple.
--
Peter Kim
This posting is provided "AS IS" with no warranties, and confers no rights.
"rajesh" <rajeshsoftdev DeleteThis @rediffmail.com> wrote in message
news:%23iO1eMVQEHA.2644@TK2MSFTNGP12.phx.gbl...
> Thanks a lot peter.
>
> We did had a look at the item no 22. But now we have following 3 queries
> regarding the same.
>
> 1. Can we set the COMPLEXITY_PENALTY factor in Analysis server?
> (Not through Create Mining model statement)
>
> 2. What is the exact significance of Minus1, Minus2 etc columns? And how
> can relate them to our Title/Store attributes?
>
> As per our understanding Minus1, Minus2 are manipulated columns to
> generate the required trend. But we are going to have 4000-5000 stores. So
> how are we going to generate these type of manipulated columns for the
> same?
>
> Looking forward to hear from you.
> Waiting for your reply.
>
> - Rajesh
>
> **********************************************************************
<font color=purple> > Sent via Fuzzy Software @ <a style='text-decoration: underline;' href="http://www.fuzzysoftware.com/</font" target="_blank">http://www.fuzzysoftware.com/</font</a>>
> Comprehensive, categorised, searchable collection of links to ASP &
> ASP.NET resources...<!-- ~MESSAGE_AFTER~ -->
>> Stay informed about: Sales Prediction 
FAQ
Profile
Private Messages
Log in
