Meme-tracking and the Dynamics of the News Cycle

by Leskovec et al. (leskovec2009, KDD2009)

The authors introduce a framework for tracking short, distinctive phrases that travel intact through on-line text (leskovec2009). Prior work identified the following two main approaches to tackle this problem:

• using propabilistic term mixtures (long range trends in general topics) [5,7,16,17,30,31]
• extracting hyperlinks and rare named entities (short information cascades) [3,14,20,23]

Their paper focuses on short units of texts, phrases and memes which are supposed to act as signatures for a topic or event (and are located between propabilitistic term mixtures and hyperlinks/rare named entities).

Method: The goal of the presented approach is creating a phrase cluster which collects phrases which are close textual variants of each other. For each cluster the authors draw the number of documents matching phrases in the cluster over time to illustrate how the attention a certain topic/meme is getting changes over time. (this is done for blogs as well as for news media to determine the time lag between these two outlets)

Evaluation: The authors tracked 1.6 million mainstream media sites and blogs over a period of three months yielding a total of 90 million articles. They identify a time lag of approximately 2.5 hours between peaks in media sites and the blogsphere and discuss global models for temporal variations and a local analysis of the peak intensity and the news/blog interactions.