Effects of Data Preprocessing on the Performance of Neural Nets for Dairy Yield Prediction and Cow Culling Classification

The impact on learning of artificial neural networks (ANN) for such factors as their architecture and the number of learning cycles has often been discussed. However, few studies have reported on the effect of data preprocessing on actual learning. According to the literature, data preprocessing could greatly improve ANNs learning ability, and may sometimes be absolutely essential, in enabling them to detect patterns in the learning data set. This project was designed to investigate the importance of data preprocessing, particularly when applying the ANN approach to classification and prediction problems. More specifically, the distribution of the output categories in the training data set was analyzed with regard to its effect on learning. For the classification case, ANNs were used to make culling decisions for dairy cows, by dividing data into one of three possible culling categories (Yes, No or Unknown). The data used in this study were previously generated by means of a prototype, fuzzy-logic based, decision-support system (DSS). For the prediction case, the ANN was to predict 305-day milk, fat and protein yield simultaneously for individual cows, using test day production information.[...