An Efficient Approach for Parameters Learning of Bayesian Network with Multiple Latent Variables Using Neural Networks and P-EM

Bayesian network with multiple latent variables (BNML) is used to model realistic problems with unobservable features, such as diagnosing diseases and preference modeling. However, EM based parameter learning for BNML is challenging if there is a large amount of intermediate results due to missing values in the training dataset. To address this issue, we propose the clustering and P-EM based method to improve the performance of parameter learning. First, an innovative layer of neural network is defined based on Recurrent Neural Network (RNN) by incorporating the structural information of BNML into the Mixture of Generative Adversarial Network (MGAN), which can reduce the number of parameters by enabling clustering in an unsupervised manner. We then propose a Parabolic acceleration of the EM (P-EM) algorithm to improve the efficiency of convergence of parameter learning. In our method, the geometry knowledge is adopted to obtain an approximation of the parameters. Experimental results show the efficiency and effectiveness of our proposed methods.