面向大规模数据集的贝叶斯网络参数学习算法

贝叶斯网络的学习可以分为结构学习和参数学习。期望最大化（EM）算法通常用于不完整数据的参数学习，但是由于EM算法计算相对复杂，存在收敛速度慢和容易局部最大化等问题，传统的EM算法难于处理大规模数据集。研究了EM算法的主要问题，采用划分数据块的方法将大规模数据集划分为小的样本集来处理，降低了EM算法的计算量，同时也提高了计算精度。实验证明，该改进的EM算法具有较高的性能。

Parameter learning for Bayesian networks with large data set

The creation of Bayesian networks can be separated into two tasks,structure learning and parameter learning.Expectation Maximization(EM) algorithm is a general method for parameter learning to incomplete data.The traditional EM algorithm has some shortcomings: it can't deal with large data sets,its convergence is slow and it easily results in local maximum.To overcome these shortcomings, large data set was divided into several small blocks and optimized in the small ones.Experiment results indicate that the improved EM algorithm has more advantages than standard EM.