Likelihood based classification in Bayesian networks
Ivan Štajduhar, Ivan Bratko · Conference on Artificial Intelligence for Applications · 2007
Learning directed probabilistic networks from data and using them for classification purposes is a well known problem. Many learning algorithms have been shown to be successful for various kinds of learning scenarios. Basically they all generate a single network from data, which is then used for classification purposes and possible domain understanding. In this paper we propose a simple method for inferring a model consisting of several Bayesian networks, each one representing data of one class. The data is divided into class subsets and from each subset a separate Bayesian network is learnt. Classification is done using prior and posterior probability distribution information in all networks. We thoroughly tested the proposed method on synthetic data and several repository datasets and compared it to other machine learning methods, to prove its effectiveness. We argue that with smaller modifications, the method can be used for learning from censored survival domains.