王杰, 严建峰, 刘晓升, 杨璐. HDP消息传递算法[J]. 微电子学与计算机, 2016, 33(3): 144-149, 154.
引用本文: 王杰, 严建峰, 刘晓升, 杨璐. HDP消息传递算法[J]. 微电子学与计算机, 2016, 33(3): 144-149, 154.
WANG Jie, YAN Jian-feng, LIU Xiao-sheng, YANG Lu. Belief Propagation Algorithm for HDP[J]. Microelectronics & Computer, 2016, 33(3): 144-149, 154.
Citation: WANG Jie, YAN Jian-feng, LIU Xiao-sheng, YANG Lu. Belief Propagation Algorithm for HDP[J]. Microelectronics & Computer, 2016, 33(3): 144-149, 154.

HDP消息传递算法

Belief Propagation Algorithm for HDP

  • 摘要: 分层狄利克雷过程(HDP)可以视为潜在狄利克雷分配(LDA)主题模型在无参方面的扩展, 解决了传统主题模型中选择主题数目的问题.消息传递算法则是从消息因子图模型的角度出发进而解决贝叶斯后验概率推断问题.提出将消息传递算法应用到分层狄利克雷过程模型上的方案, 并从最大期望算法的角度来证明该算法的收敛性, 最终从混淆度的角度将该算法与传统算法进行对比.实验结果表明消息传递算法在混淆度方面相比其他算法有明显的优势, 且收敛速度较快.

     

    Abstract: The hierarchical Dirichlet process (HDP) model is an extension of the latent Dirichlet allocation(LDA)on the aspect of non-parametric in order to solve the problem of setting number of the topics. Belief propagation algorithm is an algorithm based on the factor graph model to inference the Bayesian posterior probability. In our paper, we propose to apply the belief propagation algorithm on the HDP model, and prove the convergence of the algorithm from the view of expectation maximization algorithm. Comparing with other algorithms, the belief propagation algorithm based on HDP is better than others in accuracy measured by perplexity

     

/

返回文章
返回