I have 7 variables as inputs (4 of them are dummy) and 1 variable as output. I would like to forecast the output variable with neural network. Is there any way to choose the most relevant variables for the input of the network? Or I can select all of them as input and build/train the network?
Hi,
I think you have to do it through trial and error process since it is only seven variables. But since four of them are dummy, you may consider one-two-three-four of such variables at each step. It is often context dependent. There is no straight strategy I think.
thanks Eugene, can Automated relevance determination(ARD) help me? if yes,how can i perform that?
The pairing (input_variable_value,output_variable_value) defines the association(s) one *wants* to represent. Absent any further information it's impossible to say anything: what differentiates the ``dummy'' variables from the others, for instance? The question, as stated, simply can't be answered.
What one might want is to represent the rule {(input_variable,output_variable),(dummy_variable,output_variable)}, where one distinguishes the dummy variables, for instance, by the fact that in the association one is trying to represent, they give other values for the output variable than the input variables. And there the tiling algorithm of Mézard and Nadal can help.
Once you have developed your ANN, sensitivity analysis might help you to find out if the selected variables are ok or if you can still optimize your ANN.
- Badges
- Science topic