Approaching Multi-dimensional Classification by Using Bayesian Network Chain Classifiers

Multi-dimensional classification assigns an unseen instance to more than one class variable simultaneously. Bayesian chain classifiers have been recently proposed to address the task since they were first proposed in 2011. However, when the simplistic structure of Bayesian network is built to represent the dependency relationships among classes, the predictive performance will degrade somewhat. In this paper, we further study Bayesian chain classifiers in more explicit dependency structures for class variables by using score-based methods, that is, we learn general structures of Bayesian network by using the K2 and Hill-Climbing algorithms to represent the dependency relationships among classes. Meanwhile, we employ a fast clustering-based feature subset selection method for constructing selective Naive Bayesian network classifiers, as base classifiers. Ultimately, we develop the alternative Bayesian chain classifiers. Experimental results show that our models can be able to achieve better or at least comparable performance compared against other state-of-the-art methods, both in term of predictive performance and time complexity.