Discretisation Effects in Naive Bayesian Networks
Naive Bayesian networks are often used for classification problems that involve variables of a continuous nature. Upon capturing such variables, their value ranges are modelled as finite sets of discrete values. While the output probabilities and conclusions established from a Bayesian network are dependent of the actual discretisations used for its variables, the effects of choosing alternative discretisations are largely unknown as yet. In this paper, we study the effects of changing discretisations on the probability distributions computed from a naive Bayesian network. We demonstrate how recent insights from the research area of sensitivity analysis can be exploited for this purpose.
KeywordsBayesian Network Feature Variable Sensitivity Function Class Variable Joint Probability Distribution
Unable to display preview. Download preview PDF.
- 1.Dougherty, J., Kohavi, R., Sahami, M.: Supervised and unsupervised discretization of continuous features. In: Russel, S.J. (ed.) Proceedings of the 12th International Conference on Machine Learning, pp. 194–202. Morgan Kaufmann, CA (1995)Google Scholar
- 6.Kjærulff, U., van der Gaag, L.C.: Making sensitivity analysis computationally efficient. In: Boutilier, C., Goldszmidt, M. (eds.) Proceedings of the 16th Conference on Uncertainty in Artificial Intelligence, pp. 317–325. Morgan Kaufmann, CA (2000)Google Scholar
- 8.Dawson-Saunders, B., Trapp, R.G.: Basic & Clinical Biostatistics. McGraw-Hill, NY (2001)Google Scholar