On Early-Stage Debunking Rumors on Twitter: Leveraging the Wisdom of Weak Learners

Abstract

Recently a lot of progress has been made in rumor modeling and rumor detection for micro-blogging streams. However, existing automated methods do not perform very well for early rumor detection, which is crucial in many settings, e.g., in crisis situations. One reason for this is that aggregated rumor features such as propagation features, which work well on the long run, are - due to their accumulating characteristic - not very helpful in the early phase of a rumor. In this work, we present an approach for early rumor detection, which leverages Convolutional Neural Networks for learning the hidden representations of individual rumor-related tweets to gain insights on the credibility of each tweets. We then aggregate the predictions from the very beginning of a rumor to obtain the overall event credits (so-called wisdom), and finally combine it with a time series based rumor classification model. Our extensive experiments show a clearly improved classification performance within the critical very first hours of a rumor. For a better understanding, we also conduct an extensive feature evaluation that emphasized on the early stage and shows that the low-level credibility has best predictability at all phases of the rumor lifetime.

Appendices

Appendix A Time Period of an Event

The time period of a rumor event is hard to define. One reason is a rumor may be created for a long time and kept existing on Twitter, but it did not attract the crowd’s attention. However it can be triggered by other events after a uncertain time and suddenly spreads as a bursty event. E.g., a rumor⁷ claimed that Robert Byrd was member of KKK. This rumor has been circulating in Twitter for a while. As shown in Fig. 6(a) that almost every day there were several tweets talking about this rumor. But this rumor was triggered by a picture about Robert Byrd kissing Hillary Clinton in 2016⁸ and Twitter users suddenly noticed this rumor and it was bursted. And what we are really interested in is the tweets which are posted in hours around the bursty peak. We defined the hour with the most tweets’ volume as $t_{max}$ and we want to detect the rumor event as soon as possible before its burst, so we define the time of the first tweet before $t_{max}$ within 48 h as the beginning of this rumor event, marked as $t_{0}$. And the end time of the event is defined as $t_{end}=t_0+48$. We show the tweet volumes in Fig. 6 of the above rumor example.

Fig. 6.

tweet volume of the rumor event of Robert Byrd at full scale and after selected time period

Appendix B Full FeaturesTime Period of an Event

See Table 6.

Table 6. Features of Time Series Rumor Detection Model