Many web services, such as Twitter and Google, provide a list of their most popular terms, called a trending topics list, in descending order of popularity ranking. The changes in people’s interest in a specific trending topic are reflected in the changes of its popularity rank (up, down, and unchanged). This paper analyses the nature of trending topics and proposes a temporal modelling framework for predicting rank change of trending topics using historical rank data. Historical rank data show that almost 70% of trending topics tend to disappear and reappear later. Therefore it is important to reflect this phenomenon in the prediction model, which is related to handling missing value and window size. Missing value handling approach was selected by using expectation maximization. An optimal window size is selected based on the minimum length of topic disappearance in the same topic but with a different context. We examined our approach with four machine-learning techniques using the U.S. twitter trending topics collected from 30th June 2012 to 30th June 2014. Our model achieved the highest prediction accuracy (94.01 %) with C4.5 decision tree algorithm.
展开▼
