features selection problem..

dear all,

I have a training set matrix consists of 3 rows and 1000 columns and the target output matrix is 1 row and 1000 column, each value of a certain target column should be one of the three values in the certain training set column..

train=[g11 g12 g13.... ; g21 g22 g23....; g31 g332 g33....]; (3*1000) and the target(for example) is target=[g12 g22 g31........]; (1*1000)

the selection of target values depends on the reference values i have (i can use that reference only in training stage not testing!).. so i want to design a model to train that information and get the desire output in testing stage without using that reference values..

so, how do i design like this model matlab? i used ANN curve fitting and results were not so good..

any other idea you may advice me?

looking forward to hear from you.. best regards,

sam

[Greg Heath]

What, exactly do mean by reference value?

How many are there?

Where do they come from?

How are they appled to get the output value?

How about a simple example with size(p) = [3 4] and t = [1 4]?

Hope this helps

Greg

Answer 1

[Greg Heath]

What you are asking is not, in general, feasible.

How many reference values are used for each decision?

Assume there are two reference values and three original inputs. Then what you are asking for is equivalent to training with a 5-dimensional input and testing with only three of the inputs.

Hope this helps.

Greg

Answer 2

[wissam]

guys any help pls?

Answer 3

[Keshav Dev Singh]

hi Wissam, I was also facing the same problem in my case as I had a huge training data set to match with target. I found that for such type of problems the correlation coeff. are the best way for feature selection. You can find out its mathematical formula and try it.

Hope this helps... Keshav Dev Singh IIT Bombay

[wissam]

hi keshav,

thx for the comment..

is it possible to get such function in matlab? i am not using mathematica..

can you provide me with more details regarding that.. i mean if its possible..

thx..

wissam

[Keshav Dev Singh]

hi Wissam, try "R = corrcoef(x,y)" in matlab

--

KDSingh