摘要
概率潜在语义分析(PLSA)模型用期望最大化(EM)算法进行参数训练,由于算法参数的随机初始化,致使聚类的效果过度拟合且过分依赖于参数初始值。将潜在语义分析(LSA)模型参数概率化,用以初始化概率潜在语义分析模型的参数,得到的改进算法有效解决了参数随机初始化问题。经实验验证,所提出的方法对文本聚类的归一化互信息(NM I)和准确度都有明显提高。
Trained by the Expectation Maximization (EM) algorithm, whose model parameters are randomly initialized, the performance of Probabilistic Latent Semantic Analysis (PLSA) model is quite dependent on the initialization of the model, and the result of iteration is not a global maximum, but a local one. The authors derived probabilities from Latent Semantic Analysis (LSA), and then used it to initialize the parameters of PLSA model in documents clustering. The improved PLSA could effectively solve the puzzle of random initializing of EM. It is shown that the improved algorithm has a distinct improvement in Normalized Mutual Information (NMI) and accuracy.
出处
《计算机应用》
CSCD
北大核心
2011年第3期674-676,693,共4页
journal of Computer Applications
基金
中国博士后科学基金资助项目(20070420711)
重庆市科委基金资助项目(2008BB2191)
关键词
文本聚类
概率潜在语义分析
参数初始化
潜在语义分析
document clustering
Probabilistic Latent Semantic Analysis (PLSA)
parameter initialization
Latent Semantic Analysis (LSA)