The Impact of a Crisis Event on Predicting Social Media Virality

Understanding why specific information pieces become viral in online social networks is a fundamental step in governing social media platforms. While previous studies explored which user and network properties enable information virality, little is known about how offline crisis events impact these features. Here we investigate to what extent a crisis event impacts virality prediction models. We analyze Twitter data before and after the 2022 Russian invasion of Ukraine as a case study. We train and test statistical learning models on data before and after the invasion, evaluating which features related to content (e.g., existence of an URL), the user (e.g., number of previous tweets), or the network (e.g., degree centrality) are most relevant to predict virality. Tweets with more than 100 retweets are considered viral. We observe that the crisis event affects virality prediction: predictive accuracy is reduced when a model trained on data before the invasion is evaluated on data after the invasion—instead of being evaluated on data before the invasion. We observe that the crisis event leads to an increased fraction of viral tweets by users with fewer followers after the invasion. Moreover, we observe that, after the invasion, tweets containing URLs are more likely to become viral, implying an increased interest in website links, photos, and videos. Overall, our study reveals a shift in the importance of features underlying virality as a result of an offline crisis event.