Discriminative Learning of Bayesian Network Classifiers

F. Pernkopf (Austria)


Bayesian networks, Discriminative learning.


Recently, methods for discriminative learning of Bayesian networks used for classification, i.e. learning the structure and/or parameters by optimizing the class conditional prob ability directly, have been proposed. In this paper, we use a simple order-based greedy algorithm for learning a dis criminative network structure1 . First, we establish an or dering of the features according to the information for clas sification. Given this ordering, we can find the structure consistent with this ordering in polynomial time. We intro duce a new information theoretic score to learn the struc ture of a Bayesian network from an ordering. Furthermore, we provide a heuristic method for subsequent pruning of the learned network structure. This reduces the number of parameters and the performance may even improve due to overfitting effects, especially when the sample size for learning is small. Experiments have been performed on 25 data sets from the UCI repository. The experiments suggest that the discrimi native structure found by our algorithm outperforms on av erage other generative and discriminative structure learning approaches.

Important Links:

Go Back