Is there a solution to identify the most important independent variables after running a neural network (with a possible effect of each independent variable on the target variable if it is possible) like in logistique regression for example.
Thanks a lot
not really - how would you determine the importance of a variable based on the weights? One could, of course, use the outgoing weights from each input as an indicator: if all are zero, the variable clearly does not matter but in all other cases it does not tell you much.
You could do an interval analysis of the entire network to determine how ranges of input values affect the outputs but, again, that will only tell you parts of the story.
You can use wrapper methods, such as feature elimination to estimate the importance of variables for a given dataset, though.
Does that help?
Thank you Michael for those explanations. Backward Stepwise Elimination seems to be a good method. I know that in SPSS the contribution of inputs in the perdiction model is determined by sensitivity analysis. Can we do the same with Knime?