This module produces a number for each input variable called a contribution factor that is a rough measure of the importance of that variable in predicting the network's output, relative to the other input variables in the same network. The higher the number, the more the variable is contributing to the prediction or classification. You cannot compare contribution factors of different networks. This module works for backpropagation networks only. However, the genetic adaptive versions of the PNN and GRNN networks compute an individual smoothing factor for each input which may be thought of as a sensitivity factor for the input. The higher the value of the smoothing factor, the more impact that input has on predicting the output. Refer to PNN Learning and GRNN Learning for details. GMDH networks also indicate which inputs have more impact on predicting the output. Refer to GMDH Learning for details.
The contribution factor is developed from an analysis of the weights of the trained neural network.
You can use the contribution factor to decide which variables to remove from a network in order to simplify it and possibly cause it to train faster. You should probably only do this when your number of inputs climbs over 100 or so.
Caution: The value of a contribution factor should not to be considered "gospel" when deciding whether to include a variable in a network. Neural nets are capable of finding patterns among variables when none of the variables themselves are highly correlated to the answers. Obviously, if a certain variable is highly correlated with the answer, the variable will have a high contribution factor.
The contribution factor is also affected by the number of input variables. For example, if you include more than 60 to 80 input variables, sometimes the contribution factors get very close to each other and you can't differentiate among the variables.
The contribution factor graph displays bars which correspond to the minimum and maximum values in the input data.
Use the File Menu to select a different network to graph. The View option allows you to display a bar graph of contribution factors, display a sorted list of contribution factors in ascending or descending order with the corresponding variable name, or display an unsorted list of variable names with their corresponding contribution factors. Selecting the Export Contributions to Disk option will write out a file with the problem name and a .CO1 extension into the same directory which contains the problem files. This will be a spreadsheet file which you can view in the Datagrid or your spreadsheet program. When viewing a Weight List, the Copy from Weight List Option allows you to copy the weight list to a Windows clipboard for use in other applications.