摘要
主题模型是挖掘微博潜在主题的重要工具.然而,现有的主题模型多由Latent Dirichlet Allocation(LDA)派生,它需要用户预先指定主题数目.为了自动挖掘微博主题,作者提出了一个基于分层Dirichlet过程(Hierarchical Dirichlet Process,HDP)的非参数贝叶斯模型MB-HDP.首先,针对微博应用场景,假设消息是不可交换的;接着,利用微博的时间信息、用户兴趣以及话题标签,聚合主题相关的消息以解决微博短文本的数据稀疏问题;然后,扩展Chinese Restaurant Franchise(CRF)对微博数据进行主题建模;最后,设计一个相应的Markov Chain Monte Carlo(MCMC)采样方法,推导MB-HDP模型的分布参数.实验表明,在生成主题质量、内容困惑度和模型复杂度等指标上,MB-HDP模型明显优于LDA和HDP两种模型.
Topic models have become important tools to mine latent topics from microblogs.However,most existing models are derived from Latent Dirichlet Allocation（LDA）and require a pre-determined number of topics.In order to mine topics from microblogs automatically,we propose a hierarchical Bayesian nonparametric model named MicroBlog-Hierarchical Dirichlet Process（MB-HDP）.Firstly,our model assumes non-exchangeability of data which is suitable for the microblog application.Secondly,to tackle the sparsity problem caused by the short tweets,the temporal information,user＇s interests,and semantic #hashtags are integrated to aggregate topic-related tweets into lengthy pseudo-documents.Thirdly,the Chinese Restaurant Franchise（CRF）extension is adopted in modeling topics.Finally,we present a Markov Chain Monte Carlo（MCMC）sampling for posterior inference in the MB-HDP.Experimental results show that the MB-HDP clearly outperformed both LDA and HDP from three different perspectives：the quality of generated latent topics,the perplexity of held-out content and the model complexity.
出处
《计算机学报》
EI
CSCD
北大核心
2015年第7期1408-1419,共12页
Chinese Journal of Computers
基金
国家自然科学基金(61033010
61272065
61472453
U1401256)
广东省自然科学基金(S2011020001182
S2012010009311)
广东省科技计划项目(2011B040200007
2011B031700004
2012A010701013)资助~~
