Application of the tree augmented naive Bayes network to classification and forecasting
Adelina Tang · The University of Queensland · 2005
The thesis commences with a review of basic Static Bayesian Networks (SBNs) and describesnsome of the methods of probabilistic inference from SBNs, including Pearl's causal treenmethodology and the more general NP-complete clique tree methodology due to Lauritzennand Speigelhalter. The use of SBNs in pattern classification will then be described withnspecial consideration given to the use of elementary forms of SBNs including the NaivenBayes and the Tree Augmented Naїve Bayes (TAN) classifiers. The possibility of applyingnboosting and other ensemble methods to these elementary forms in order to obtain superiornperformance will then be explored. The discussion will conclude by investigating thensuitability of correlation measures in computing the TAN as an initial step in adapting it tonsolve the forecasting problem. Attention will then be turned to the use of Bayesian Networksn(BN) in forecasting. Numerous techniques have been developed to create accurate forecastingnmodels and the BN approach, with a time element incorporated, is among the more successfulnof these. The time element will be introduced by means of a series of SBNs, acting as time-slicesnto create a Dynamic Bayesian Network (DBN), to solve the forecasting problem. Thensuitably modified TAN combined with the Pearl causal tree, now called the TAN-PearlnNetwork (TPN), will form the basis of the DBN. The objective of the forecasting problemnwill be to minimize overall error, computed from the differences between the computednbeliefs at each time-slice, and those of the actual events. The final major investigationnconcerns the application of boosting to regression. It draws upon the parallel between time-slicesnin a DBN and instances in regression analysis and it will be shown that the accuracy ofnthe time-slices, and therefore that of the DBN as a whole, can be improved through boosting.n