Applies to: NeuralTools 5.x and Up

 

NeuralTools Predicts Only the Majority Class in a Binary Dependent Variable

Question:
My dependent variable is all 1’s and 0’s, with many more 1’s than 0’s. After training, NeuralTools predicts all 1’s correctly but gets all 0’s wrong. What’s happening?

Answer:
NeuralTools is encountering an unbalanced dataset. Because there are too few 0 cases, the network “learns” that most cases are 1 and predicts 1 for every instance. While this maximizes overall accuracy, the predictions aren’t useful because the minority class is ignored.

---

Solutions

1. Use Oversampling

• NeuralTools includes an example illustrating oversampling to balance classes:
  Help → Example Spreadsheets → Other Category Prediction Examples → Advertising Responses - Oversampling

• This technique increases the representation of the minority class in training, improving the network’s ability to predict both classes.

2. Try a Best Net Search with an MLF Network

• PNN networks predict by interpolation and struggle with highly unbalanced training sets.

• MLF networks can find general patterns, making it more likely to achieve good predictions even with unbalanced data.

• Note: MLF networks do not return probabilities of predictions.

Tip: Always examine your dependent variable’s distribution before training. If one class dominates, consider resampling or choosing a network type better suited to handle imbalance.