Save (Multilayer Perceptron)
The Save tab is used to save predictions as variables in the dataset.
- Save predicted value or category for each dependent variable. This saves the predicted value for scale-dependent variables and the predicted category for categorical dependent variables.
- Save predicted pseudo-probability or category for each dependent variable. This saves the predicted pseudo-probabilities for categorical dependent variables. A separate variable is saved for each of the first n categories, where n is specified in the Categories to Save column.
Names of Saved Variables. Automatic name generation ensures that you keep all of your work. Custom names allow you to discard/replace results from previous runs without first deleting the saved variables in the Data Editor.
Probabilities and Pseudo-Probabilities
Categorical dependent variables with softmax activation and cross-entropy error will have a predicted value for each category, where each predicted value is the probability that the case belongs to the category.
Categorical dependent variables with sum-of-squares error will have a predicted value for each category, but the predicted values cannot be interpreted as probabilities. The procedure saves these predicted pseudo-probabilities even if any are less than 0 or greater than 1, or the sum for a given dependent variable is not 1.
The ROC, cumulative gains, and lift charts (see Output (Multilayer Perceptron) ) are created based on pseudo-probabilities. In the event that any of the pseudo-probabilities are less than 0 or greater than 1, or the sum for a given variable is not 1, they are first rescaled to be between 0 and 1 and to sum to 1. Pseudo-probabilities are rescaled by dividing by their sum. For example, if a case has predicted pseudo-probabilities of 0.50, 0.60, and 0.40 for a three-category dependent variable, then each pseudo-probability is divided by the sum 1.50 to get 0.33, 0.40, and 0.27.
If any of the pseudo-probabilities are negative, then the absolute value of the lowest is added to all pseudo-probabilities before the above rescaling. For example, if the pseudo-probabilities are -0.30, 0.50, and 1.30, then first add 0.30 to each value to get 0.00, 0.80, and 1.60. Next, divide each new value by the sum 2.40 to get 0.00, 0.33, and 0.67.
How To Save Variables to the Active Dataset
This feature requires the Neural Networks option.
- From the menus choose:
- In the Multilayer Perceptron dialog box, click the Save tab.