一种学习稀疏BN最优结构的改进K均值分块学习算法

At present, the traditional structure learning algorithm of Bayesian networks (BN) shows the problem of excessive computational burden and difficulty in obtaining the desired accuracy in a reasonable time when processing high-dimensional data. In order to learn the optimal structure of sparse BN under high-dimensional data, this paper proposes a block learning algorithm with improved K-means algorithm for learning sparse BN optimal structure. The algorithm adopts the strategy of divide and conquer. Firstly, we use mutual information as the distance between nodes, and the improved K-means algorithm with mutual information is used to block the network. Secondly, the MMPC algorithm is used to obtain the skeleton of the whole network. According to the skeleton, the possible connection directions of all edges between the blocks are found, so that all possible graph structures are found; after that, structural learning is performed sequentially for all possible graph structures; finally, the best BN is found by using scoring function. Experiments show that compared with the existing block structure learning algorithm, the proposed algorithm not only learns the optimal structure of the network, but also improves the learning speed definitely. Compared with the non-blocking classical structure learning algorithm, the learning speed of the algorithm proposed in this paper is greatly improved on the basis of ensuring accuracy, which solves the problem that the traditional algorithms cannot process high-dimensional data in a reasonable time.