Using Neural Networks with Different Target Settings to Predict Maternal Health Risks

Abstract

This paper proposes a technique for training neural networks with multiple outputs using different target settings. Three neural networks with the same configuration are set up with different targets. They are trained using the same data to predict different outputs, which are then combined to produce the final result. Maternal health risk dataset from the UC Irvine machine learning repository is used to test the proposed technique. This technique can achieve better accuracy than an ensemble neural network and stacking neural network trained with the original targets. In addition, the proposed technique can give better accuracy when combined with cascade generalization.