基于无限逆狄利克雷混合模型的变分学习

Variational Learning Based on Infinite Inverted Dirichlet Mixture Model

近期的研究表明,有限逆狄利克雷混合模型是一种建模非高斯数据的重要的模型。然而,它存在参数估计及模型选择困难的问题。利用常用的EM算法无法对其进行准确地估计参数及选择最佳的混合分量数。因此,论文研究无限逆狄利克雷混合模型,提出一种变分近似推理算法对其进行学习。该算法能够同时解决这两个问题。为了验证算法的有效性,论文在人工数据集上进行了大量的实验,实验结果表明利用变分贝叶斯推理来估计混合无限逆狄利克雷分布是一种非常有效的方法。

Recent studies have shown that finite inverse Dirichlet mixture model is an important model for modeling non Gauss data.However,it has the problem of parameter estimation and model selection.The EM algorithm can not be used to accurately estimate the parameters and select the optimal number of mixture components.Therefore,this paper studies the infinite inverse Dirichlet mixture model,presents a novel variational approximate inference algorithm for learning.The algorithm can solve these two problems at the same time.In order to verify the effectiveness of the algorithm,this paper carries out experiments on artificial data sets.Experimental results show that the variational Bayesian inference to estimate mixed infinite inverse Dirichlet distribution is a very effective method.