Re: Question about Decision Trees and Neural Networks

Tech-Archive recommends: Repair Windows Errors & Optimize Windows Performance

From what I can see here, you do have certain records with negative amout,
so it won't be a big surprise that Neural Network output some negative
predictions. As to the classification accuracy, there are no guranttee that
one algorithm will do better than another. You can try to tune the
performance of the nural network by changing the following parameters:

Default

MAXIMUM_STATES 100

Sample_SIZE 10000

From what I can see, your attributes have many states, you you can increase
the MAXIMUM_STATES. Also, Nueral Network may need a very large training data
set to achieve good performance on you data, you might also what to
dramatically increase that number. Another suggestion is not to use
Account-ID-type attributes as anything besides a key.

Good luck,
--
-Yimin Wu
SQL Server Data Mining
[This posting is provided "AS IS" with no warranties, and confers no
rights.]


<anonymous_user@xxxxxxxxxxxxxxxxxxxxxxx> wrote in message
news:3683ca13-0766-49e2-b07b-0b678962098a@xxxxxxxxxxxxxxxxxxxxxxx
Here's some sample of data:

ID Account Org Dim01 Dim02 Avg Amount
--------------------------------------------------------------
11 6812 20556400 7760 # 492 255
12 6812 20556400 7760 # 492 304,28
13 1510 20999995 # # 1772 2350
14 3010 20305000 7100 # 1815 -2350
17 3629 45420000 3280 920490 5742 5954,5
19 7612 10024000 0040 177459 250 300
22 6430 50110000 0540 # 1652 4065,1
23 1510 60999998 # # 4830 3523
26 6431 26301000 4710 # 3408 75
29 3131 45230000 3410 # 4371 -3966

Account, Org, Dim01, Dim02 is discrete text attributes.
Avg and Amount is long Continous attributes.
Avg is the avarage amount of the combination of Account, Org, Dim01,
Dim02:s Amounts.
I'm trying to predict Amount.




As to the second question, it will be helpful if you can provide a small
sample of your data. From what you post, it seems that all your data are
numerical (i.e., numbers), because you got a average-amount attribute
already. If my guess were right, it might be of help to change the data
type
of all attributes to continuous int/double for the Neural Network.

Good luck,
--
-Yimin Wu
SQL Server Data Mining
[This posting is provided "AS IS" with no warranties, and confers no
rights.]



"Lars E" <lae_1980@xxxxxxxxxxx> wrote in message
news:5cc24cd0-1e44-4342-b6de-0d94ab75b438@xxxxxxxxxxxxxxxxxxxxxxx
I have some accounting data, with some transaction attributes and
amounts.
I'm using Decision Trees to try and predict the next month's amount for
certain combinations of attributes.

I've tried two different structures for the model:

A: one with 9 discrete text input attributes.
B: And another with the same 9 attributes + a avarage Amount for all
combinations of the nine attribute for every transaction.


When i've processed them and look in the dependency network, it says
that
the strongest link for the structure A is attribute "1".
And for the second its the avarage-Amount attribute.
Okey, that seems fine, but the second strongest link in structure B is
attribute "2".

Shouldn't it be attribute 1 like in structure A?



Second question, if I run the same data in a Neural Network model, the
prediction becomes much worst then the decision tree.
I get many predictions that are negative values even though all
training
data contains positiv values.
The StDev becomes the same for every row also..
What am I doing wrong with that one. I have alot of transactions and a
read somewhere that a Neural Network should work better than a decision
tree in a case similar to mine.
The score in the "Lift chart" for the Neural Network model becomes 0,00
and for Decision Trees with the same data I get around 110.




.

Relevant Pages

  • Question about Decision Trees and Neural Networks
    ... And another with the same 9 attributes + a avarage Amount for all combinations of the nine attribute for every transaction. ... I have alot of transactions and a read somewhere that a Neural Network should work better than a decision tree in a case similar to mine. ... The score in the "Lift chart" for the Neural Network model becomes 0,00 and for Decision Trees with the same data I get around 110. ...
    (microsoft.public.sqlserver.datamining)
  • Re: Question about Decision Trees and Neural Networks
    ... Avg is the avarage amount of the combination of Account, Org, Dim01, Dim02:s Amounts. ... Second question, if I run the same data in a Neural Network model, the ... and for Decision Trees with the same data I get around 110. ...
    (microsoft.public.sqlserver.datamining)
  • Re: Speed of SQL 2005 decision tree vs one coded in C#?
    ... I am easily executing several hundred thousand of these singleton queries per hour, but I'd like to execute millions per hour. ... Batch predictions have typically much better performance than singleton ... I am using decision tree and neural network algorithms. ...
    (microsoft.public.sqlserver.datamining)
  • Re: Speed of SQL 2005 decision tree vs one coded in C#?
    ... Batch predictions have typically much better performance than singleton ... I am using decision tree and neural network algorithms. ... I have a C# program which opens a connection to AS, ...
    (microsoft.public.sqlserver.datamining)
  • Re: the best forecasting method
    ... Neural Network ... position of a pendulum versus time, then nonlinear regression would be ... customer categories based on demographic information, then decision trees or ... http://www.dtreg.com (decision tree modeling) ...
    (sci.stat.math)