Fast Identification of Topic Burst Patterns Based on Temporal Clustering

Temporal text mining is widely used in summarization and tracking of evolutionary topic trends. In online collaborative systems like Wikipedia, edit history of each article is stored as revisions. Topics of articles or categories grow and fade over time and retain evolutionary information in edit history. This paper studies a particular temporal text mining task: quickly finding burst patterns of topics from phrases extracted from edit history of Wikipedia articles. We first extract several candidate phrases from edit history by specific features and build time series with edit frequency. Temporal clustering of burst patterns of phrases reveals bursts of topics. However, distance measure for temporal clustering, such as dynamic time warping (DTW), is often costly. In this paper, we propose segmented DTW which decomposes time series into proper segments and computes DTW distance within segments separately. Our segmented DTW shows reasonable speed up over DTW, while the proposed method can identify interesting evolutionary topic burst patterns effectively. Research so far can be applied in domains like trend tracking, temporal relatedness of phrases and popular topic discovery.