Online Learning of Bayesian Network Parameters

The paper introduces a novel online learning algorithm of Bayesian network parameters. Inspired by maximum likelihood estimation, we modify Voting EM algorithm, which is an online parameter learning method proposed by Cohen, et al, to gain better results, making the parameters of the algorithm easier to be specified. It adjusts the learning rate by changing the weights of samples according to the time they arrived at. And that means it is more suitable to be applied in practice. We demonstrate the performance of the proposed method through the comparison with Voting EM. The result confirms that the proposed method is easier to get a better estimate of the Bayesian network parameters, and also adapts to the new parameters quickly. Further more, the accuracy of the estimation is improved.