Predicting the Occurrence of Life Events from User's Tweet History

This paper addresses the new problem of predicting the occurrence of Twitter user's life events using word occurrence tendencies in user's tweet histories. Many previous studies have addressed public event prediction and life event extraction on Twitter. Most of the methods for these two problems use tweets that explicitly refer to event occurrence. However, users who will experience a life event are unlikely to post tweets that reference the occurrence of the life event explicitly; thus, existing methods to find tweets that refer to event occurrence explicitly are not applicable to life event prediction. Therefore, we assume that users who will experience a life event tend to post tweets that refer to life event occurrence implicitly and propose a method to identify such a tendency to predict life events. First, we extract users who experienced a specific life event and collect their past tweets to identify features that implicitly indicate the life event occurrence. We use the word occurrence tendency in such tweets as training data to construct a life event prediction model. We chose five life events, that is, "Giving birth," "Getting a job offer," "Leaving the hospital," "Pregnancy," and "Marriage," and assessed the prediction performance of the proposed method for each event. Experimental results demonstrate that the proposed method outperformed a baseline method for all selected life events except "Leaving the hospital" and achieved the highest prediction accuracy for "Giving birth." We suppose that this event resulted in the highest prediction accuracy because all users who gave birth experienced pregnancy and common features appeared in their tweets over a long period.