Maximizing the reliability of two-state automaton for burst feature detection in news streams
Gang Du, Jun Hai Guo, Weiran Xu, Zhen Yang · 2010
The capture of temporal dynamics of news streams has drawn increasing attentions in recent sequential data mining works. Most of them are based on the intuition that a “burst” of a topic is signaled by a growth of relevant words in a high intensity during a period of time. Such “burst features” can be efficiently identified by Kleinberg's two-state automaton model. The resolution is an important parameter of the model. It affects the reliability of the results greatly. This paper maximizes the reliability of the results by estimating adaptive resolution for each word with EM algorithm. Experiments with the public news corpora prove that the unified resolution is a bottleneck of the performance, and the results with word-adaptive resolutions approximate to the maximum reliability well.