Discovering bursty topics from news stream with a unified statistical model

作     者：Qi, Xiang Huang, Yu Liu, Xiaoyan Fu, Kun Wang, Hongqi 

作者机构：Key Laboratory of Technology in Geo-Spatial Information Processing and Application System Institute of Electronics University of Chinese Academy of Sciences Beijing China 

出 版 物：《Journal of Computational Information Systems》 (J. Comput. Inf. Syst.)

年 卷 期：2014年第10卷第18期

页      面：7825-7832页

核心收录：

主　　题：Bayesian networks 

摘      要：This paper focuses on discovering bursty topics from news stream. Previous work usually apply Kleinberg s modeling of burst to topics estimated by a topic model such as Latent Dirichlet Allocation (LDA) and Dynamic Topic Model (DTM). However, Kleinberg s model is originally proposed for the burst of keywords, the frequency counts it models are not proper to describe the burst states of topics, leading to some unwanted results. A more reasonable way is to model the influence burst states put on each document s topic distribution. Considering this, we propose a unified statistical model that takes the burst states as markov latent variables that influence the topic allocation of documents. We derive a Gibbs sampling algorithm for the proposal. Experiment results confirm our model s advantages both qualitatively and quantitatively.