Detecting evolutionary stages of events on social media: A graph-kernel-based approach

Detecting the evolutionary stages of social media events such as Twitter and Sina Weibo is beneficial for enterprises and governments to take necessary actions before emergent phenomena become uncontrollable. Prior work on event extraction from microblogs mostly focused on extracting event summary. However, the evolutionary stages of an event can provide more details about the evolution of the event, which are more critical to predict the future trend of the event. On the other hand, many events have a lifecycle-like evolutionary property, i.e., beginning, developing, climax, descending, and disappearing. Such a feature is helpful to detect the evolutionary stages of events. Thus, we propose a graph-based approach to represent and extract the evolutionary stages of events from microblogs based on this consideration. The contributions of this study are threefold. First, differing from existing methods that use a keyword set or a microblog set to represent an event, we propose a Keyword Popularity Information Graph (KPIG) to represent the keywords and the statistical information of events using a graph. With this mechanism, we can capture both the textual information and the statistical information of an event. Second, based on the KPIG graph, we present a graph-kernel-based approach to measure the similarity among events. Third, we conduct extensive experiments on a real dataset and compare our proposal with several competitor algorithms. The results show that our approach outperforms other competitors in terms of various metrics. (C) 2021 Published by Elsevier B.V.