Interpretation of Hidden Node Methodology with Network Accuracy

Bayesian networks are constructed under a conditional independency assumption. This assumption however does not necessarily hold in practice and may lead to loss of accuracy. We previously proposed a hidden node methodology whereby Bayesian networks are adapted by the addition of hidden nodes to model the data dependencies more accurately. Empirical results in a computer vision application to classify and count the neural cell automatically showed that a modified network with two hidden nodes achieved significantly better performance with an average prediction accuracy of 83.9% compared to 59.31% achieved by the original network. In this paper we justify the improvement of performance by examining the changes in network accuracy using four network accuracy measurements; the Euclidean accuracy, the Cosine accuracy, the Jensen-Shannon accuracy and the MDL score. Our results consistently show that the network accuracy improves by introducing hidden nodes. Consequently, we were able to verify that the hidden node methodology helps to improve network accuracy and contribute to the improvement of prediction accuracy.