Rumor Detection over Varying Time Windows

Abstract

This study determines the major difference between rumors and non-rumors and explores rumor classification performance levels over varying time windows-from the first three days to nearly two months. A comprehensive set of user, structural, linguistic, and temporal features was examined and their relative strength was compared from near-complete date of Twitter. Our contribution is at providing deep insight into the cumulative spreading patterns of rumors over time as well as at tracking the precise changes in predictive powers across rumor features. Statistical analysis finds that structural and temporal features distinguish rumors from non-rumors over a long-term window, yet they are not available during the initial propagation phase. In contrast, user and linguistic features are readily available and act as a good indicator during the initial propagation phase. Based on these findings, we suggest a new rumor classification algorithm that achieves competitive accuracy over both short and long time windows. These findings provide new insights for explaining rumor mechanism theories and for identifying features of early rumor detection.

Fig 1. Log-log scale of the CCDF.

Complementary Cumulative Distribution Function (CCDF) of aggregated user characteristics for rumor and non-rumor events in the 56-day observation.

Fig 2. Diffusion network examples.

The network visualization in (a) shows that rumors involve a larger fraction of singletons and smaller communities, resulting in a sporadic diffusion pattern. In contrast, the diffusion network of a non-rumor event in (b) is highly connected, forming a giant connected component and a smaller fraction of singletons. Edge colors represent the relative influence of the spreader and recipient, such that red (blue) means information propagated from a lower-degree (higher-degree) spreader to a higher-degree (lower-degree) recipient.

Fig 3. Samples of extracted time series.

The time series are extracted from 56-day observation period(x-axis = days; y-axis = number of tweets). Rumors typically have longer life spans and more fluctuations.

Fig 4. Correlogram among the linguistic features.

These plots show correlations among linguistic features with significance for rumor events. Colors of circles represent correlation coefficients, where the dark blue (red) color indicates coefficients of 1 (-1). Size of circles represent the absolute values of correlation coefficients. Blank or no circle indicates non-significant cases where p-values ≥ 0.05. From (b), strong negation is expressed by a positive correlation between scores of ‘negation’ and ‘certain’. Interestingly, the ‘assent’ category increases for rumors, which indicates that users who confirmed the rumor also appeared as time passed.

Fig 5. Comparison of the strength of the features in determining rumors.

Total and User+Linguistic are the newly proposed rumor classification algorithms in this study.