Detect Rumors Using Time Series of Social Context Information on Microblogging Websites

Contributions

• DSTS model — a temporal feature representation that captures both absolute feature values and slopes (trends) between consecutive time intervals
• Multimodal feature engineering — 26+ content-based, user-based, and diffusion-based features with time-aware aggregation
• Early detection analysis — shows the method reaches baseline accuracy 10–50 hours faster than competing methods
• Cross-platform validation — evaluation on both Twitter (207 events) and Sina Weibo (922 events) with verified rumor labels

Method

The approach discretizes event lifecycles into N fixed time intervals (N=50 in experiments). For each interval t, features are aggregated and normalized using Z-score across all intervals.

DSTS representation (Equation 3):

• Absolute features: F^D_{i,t} for each interval t (content statistics, user attributes, retweet counts)
• Slopes: SD_{i,t} = (F^D_{i,t+1} − F^D_{i,t}) / Interval(E_i), capturing rate of change

This combines both the cumulative statistics and the temporal trajectory of each feature. A linear SVM classifier is trained on the combined DSTS feature vector.

Feature types:

• Content-based (14 features): LDA topic distributions, sentiment scores, URL/emoticon/punctuation percentages, first-person pronouns, hashtags
• User-based (10 features): verified status, follower/friend counts, account age, profile completeness, gender distribution
• Diffusion-based (3 features): retweet/comment counts, microblog volume

The key insight: static baselines (e.g., total retweets) miss that rumors and non-rumors follow different temporal patterns in question-mark usage, first-person pronouns, and sentiment evolution.

Results

Full lifecycle detection (Table 3):

• Twitter: 89.6% accuracy (10.0 point F1 improvement over DT; 4.4 point improvement over RF-ext)
• Sina Weibo: 84.6% accuracy (9.3 point accuracy improvement over DT)
• Outperforms non-linear baselines (RF, SVM-RBF) despite using a linear SVM, indicating DSTS features are highly predictive

Early detection (Figure 2):

• On Twitter, DSTS reaches RF's peak accuracy in ~15 hours; RF requires 25+ hours
• On Weibo, DSTS achieves similar performance at 20 hours; RF needs 70 hours

Notes

Strengths:

• Compelling temporal intuition with clear Figure 1 examples showing different time patterns in rumors
• Large-scale evaluation across two different platforms and languages
• Ablation studies (Table 3) show all feature types contribute; content alone is competitive

Weaknesses:

• No theoretical justification for why slopes specifically capture rumor signals; empirical finding only
• Limited to supervised classification on pre-labeled events; no zero-shot or cold-start scenarios
• Time interval discretization (N=50) appears fixed; sensitivity to this hyperparameter not explored