Noise Smoothing in Learning Parameters of Bayesian Network
Du R · Jisuanji fangzhen · 2009
Noise universally exists in discrete data in real life.It is important and difficult how to smooth these noises effectively.At present,the methods of reducing noise influence are to simplify the structure in the parameter learning of Bayesian network.But the regulated structure will lose some useful edges to fall network reliability.The influence of noise can not be decreased or removed from root.And the continuous learning is very difficult.An effective method of smoothing noise was proposed by combining Bayesian network with Gibbs sampling.A variable was taken as a basic unit and its corresponding datum were smoothed by using the information of Markov blanket providing.The problems above can be efficiently avoided.