Parameter Importance For Neural Network

Hi all, i am trying to see which of my input is more important when using neural network, something like [ TreeBagger(100,Xtrain,Ytrain(:,1),'method','regression','oobvarimp','on'] which allows me to see which of the parameters does the model replies on more to predict the outputs. I tried to use the wb = net.IW to get the input weight to determine which parameter is actually more important but what i don't understand is when i do that the weight contains negative value and even exceed 1 ?

Answer 1

[Greg Heath]

[ I N ] = size(input) % = ?

[ O N ] = size(target)% = ?

0. If inputs are not orthogonal and/or N is large, don't expect a clear cut definitive answer.

1. Standardize both inputs and targets to zero-mean/unit-variance using zscore

2. As a guide, obtain input rankings for a linear model using both backward and forward searches.

 help stepwisefit
 doc stepwisefit

3. My typical design goal is : Minimize the number of NN hidden nodes, H, subject to the constraint MSE <= 0.01 (then 99% of the target variance is successfully modeled). Others may differ

4. Obtain 10 I-H-O nets, trained from different initial states of the RNG, satisfying:

 a. The number of hidden nodes, H, is a minimum
 b. Subject to the constraint 
    Either 
    MSE <= 0.01
    or
    MSE is minimized but cannot achieve 0.01

5. For each of the 10 nets you will get 10 rankings from which you can make your choice

 a. Rank the inputs by the MSE obtained when each input is replaced 
    by zeros
 b. Replace the input that results in the lowest MSE with zeros.
 c. Repeat the process until all inputs are ranked.

Hope this helps.

Thank you for formally accepting my answer

Greg

[Greg Heath]

Since all inputs are standardized, replacing inputs with zeros is equivalent to removing the input.

However, if inputs are removed, the remaining inputs have to be renumbered.

Yuck.

[crixus]

Hi thanks for your reply, for now i'm doing a loop to delete columns by columns and record the accuracy when each column is remove, those columns when removed that result in accuracy being improved, i took it as noise or unimportant variables. My accuracy is measured in R2 value, not sure if this is what you meant initially.