Home → Techniques and Tips → NeuralTools → Dependent Variable is Binary 1/0 or Yes/No
Applies to: NeuralTools 5.x–7.x
My dependent variable is all 1's and 0's, with a strong preponderance of 1's and few 0's. After training, NeuralTools simply gets all the 1's correct, and all the 0's wrong. What's wrong?
Basically, NeuralTools doesn't have enough 0 cases to learn from. It has decided that most cases result in a 1, and it just predicts a 1 every time. It's right most of the time, but the predictions aren't useful.
In NeuralTools, click Help » Example Spreadsheets » Other Category Prediction Examples » Advertising Responses - Oversampling. This example illustrates oversampling to make a better balance between the 0 and 1 cases. (The example has a preponderance of 0's rather than a preponderance of 1's, but the principle is the same.)
As an alternative, you might try a Best Net Search on your original data. The search may find an MLF with good testing results. (PNN nets predict by interpolation from training data, so they don't work too well if the training set is very unbalanced, with the vast majority of the training cases falling into one of the dependent categories. MLF nets are capable of finding general patterns, so it is more likely that a good MLF net can be found with an unbalanced training set. But keep in mind MLF nets don't return probabilities of predictions.)
Last edited: 2015-09-03