Feature importance of fully connected neurons network (Relative Importance)

Dear MathWorks Community,

I have a fully connected neurons network, and I have the initial and result in weights. The network has five inputs and three neurons with one output; the bias is (+1). I kindly ask if it is possible to calculate the feature importance (Relative Importance) of the inputs; let us Call them (A, B, C, D, and E). The initial and the resulting weights are below. It is appreciated if the number of inputs and neurons be as parametric to be changed.

////initial weights

0.96000000 -0.78000000 -0.76000000 -0.42000000 -0.48000000 0.52000000

0.52000000 0.94000000 0.74000000 0.22000000 -0.06000000 0.36000000 -0.50000000

-0.90000000 -0.66000000 -0.32000000 0.02000000 0.88000000 0.10000000 -0.68000000 0.38000000

////result weights

30.28172977 -8.19280617 2.59276326 -7.30170239 -1.79862626 0.39566791

99.09610372 -3.98615863 41.15997085 24.65729897 383.02673592 41.82976873 -177.63460075

-31.48093460 -0.05714430 -0.03576784 0.01020878 -0.05643959 0.00721399 -0.13373771 31.67916521

Thanks in advance

Sherwan

Answer 1

[Vatsal]

Hi,

I understand that you want to calculate the feature importance (Relative Importance) of the inputs. Attached below is a general approach to calculate the relative feature importance:

% Define the initial and result weights 
initial_weights = [0.96, -0.78, -0.76, -0.42, -0.48, 0.52;  

                   0.52, 0.94, 0.74, 0.22, -0.06, 0.36;  

                   -0.5, -0.9, -0.66, -0.32, 0.02, 0.88]; 

result_weights = [30.28, -8.19, 2.59, -7.30, -1.79, 0.39;  

                  99.09, -3.98, 41.15, 24.65, 383.02, 41.82;  

                  -177.63, -31.48, -0.05, -0.03, 0.01, -0.05]; 

 
delta_weights = abs(result_weights - initial_weights); 
sum_changes = sum(delta_weights(:,1:num_inputs), 1); 
total_sum_changes = sum(sum_changes); 
% Calculate the relative importance 

relative_importance = sum_changes / total_sum_changes; 

I hope this helps!