SCSNet Improves Early Rumor Detection Across Unseen Events

A neural network designed to classify online rumors while discussions are still unfolding outperformed competing systems on three public benchmarks, including an event-separated evaluation of unseen PHEME news events. The Stage Contribution State Network, or SCSNet, achieved mean Macro-F1 scores of 89.34% on Twitter15, 90.86% on Twitter16 and 82.74% on PHEME. In the PHEME cross-event test, the model’s Macro-F1 exceeded the closest competing method by 1.24 percentage points. The study also found consistent improvements as the observation window increased from 10 to 30 minutes, indicating that SCSNet could extract useful evidence before a complete propagation tree had formed.

The underlying challenge is that early detection provides limited material for a judgment. A source post may have accumulated only a few replies, some of which are emotional reactions, repeated reposts or comments unrelated to the claim being assessed. Different events can also have substantially different vocabularies, participant groups and diffusion rhythms, making systems that learn event-specific patterns less reliable when applied to new incidents. SCSNet therefore does not treat every reply, and every posting time, as equally important. Instead, it estimates how much each early reply contributes to the current judgment, groups replies with similar estimated roles into contribution stages and then updates an event state progressively as those stages are processed.

Technically, the model begins with a source message, the replies visible during an observation window, parent-node relations among replies and their arrival times. A text encoder, initialized with bert-base-uncased, produces representations for the source and reply messages. SCSNet then measures two semantic differences for every reply: the difference between the reply and the source, and the difference between the reply and its parent message. The model combines these differences with arrival time and propagation depth to represent roles that may include questioning, supporting, supplementing or clarifying information. Reply time is transformed with a logarithmic function, while depth is normalized by the maximum depth in the current window. This allows the network to combine content and structural position without relying directly on the absolute speed or scale of an event’s propagation.

These elements form an organized representation containing the reply’s semantic differences from its source and parent, together with learned temporal and depth directions. A sigmoid function then converts that representation, together with normalized time and depth, into a contribution coefficient between zero and one. The coefficient determines how strongly a reply influences its assigned stage. Replies with similar coefficients are not grouped strictly by chronology; instead, a soft-assignment formula calculates their membership in a fixed number of stages, set to four in the experiments. Each stage vector is a weighted combination of its replies, with the contribution coefficients also acting as weights. This design allows repeated or weakly informative messages to have less influence while retaining the order in which distinct information states are processed.

SCSNet then uses a linear state-update mechanism rather than a recurrent or attention-heavy sequence model. A diagonal retention matrix controls how much of the previous event state survives at each step. Before the state changes, the network compares the current stage with the prediction implied by the earlier state and generates an update gate from the difference. A low gate value limits the effect of a stage estimated to add little new information; a higher value permits a more substantial update. The event state is therefore advanced sequentially from one contribution stage to the next, using only the preceding state and current stage. Because a prediction can be generated at every stage, the model can issue an early classification without waiting for the full cascade, while the state-consistency loss penalizes unnecessary changes between adjacent stages.

Training combines three objectives. Every stage contributes to classification loss, but later stages receive greater weight than earlier ones, allowing the model to learn from more complete information while still learning to classify at the first observation point. A contribution-constraint term discourages the model from assigning almost identical value to all replies, while the state-consistency term discourages unstable updates when neighboring stages contain repetitive information. Experiments used a stage-assignment temperature of 0.10, a contribution-loss weight of 0.10 and a state-consistency weight of 0.20. Five matched random seeds were used for the main, early-window and cross-event comparisons. Paired t-tests produced p-values below 0.001 for the reported comparisons across five overall metrics on all three datasets.

Ablation experiments indicated that both the representation and the update mechanism contributed to the performance. Adding contribution-based information organization alone improved accuracy by 2.94 percentage points on Twitter15, 2.80 points on Twitter16 and 3.23 points on PHEME. Removing the contribution constraint reduced PHEME Macro-F1 from 82.74% to 79.82%, while removing the update gate lowered Twitter16 F1 from 91.40% to 89.52%. Controlled comparisons used the same contribution-organized sequences but replaced the linear update with a GRU, LSTM or lightweight Transformer. SCSNet still performed better: its Macro-F1 exceeded the strongest controlled alternative by 0.71 points on Twitter15, 1.03 on Twitter16 and 1.13 on PHEME. The clearest ablation improvement appeared for unverified rumors, whose F1 rose by 7.40 points on Twitter15 and 7.53 points on Twitter16 when the full system was used.

Early-window tests showed that the advantage was most pronounced when evidence was scarce. At 10 minutes, Twitter15, Twitter16 and PHEME contained an average of only 3.8, 4.2 and 2.9 visible replies, respectively. Macro-F1 increased by 3.52 ± 0.38 points on Twitter15, 3.50 ± 0.35 points on Twitter16 and 3.63 ± 0.44 points on PHEME as the window extended from 10 to 30 minutes, while the number of visible replies and propagation depth increased and failure rates declined. The study also transferred a model trained on combined Twitter15 and Twitter16 data to 400 Twitter-COVID19 events without target-domain fine-tuning. It reached 79.26 ± 0.42% accuracy and 76.91 ± 0.48% Macro-F1 on that unseen domain, outperforming the closest competitor by 1.24 and 1.28 percentage points, although performance was lower than in the source domains.

The computational results also favored the relatively simple architecture. SCSNet contained 24.7 million parameters and required 13.2 milliseconds per event for inference. Its parameter count, computational cost, inference latency and memory use were lower than those of the comparison methods in the study. Comparisons with four recent large-language-model-based rumor detectors also placed SCSNet narrowly ahead, including an advantage of 0.33 percentage points in PHEME Macro-F1 over the strongest listed LLM-based competitor. These results do not establish that a smaller specialized network is superior to large language models in every setting: the comparison was limited to the datasets, label configurations and protocols used by the researchers. The method also cannot replace fact checking, and its benchmark results do not demonstrate the effectiveness of automated moderation in live platforms.

Several failure modes define the boundaries of the approach. Extremely sparse discussions provide too little evidence for stable stage representations, while sarcasm, jokes and other implicit pragmatic meanings can cause replies to appear more informative than they are. Emotional reactions to true news may also be assigned excessive contribution, and strong entity or location cues can interfere with the estimated semantic relationship between a reply and its source. The experiments focused on English-language social media, leaving cross-lingual differences untested. The authors propose uncertainty calibration, pragmatic-function recognition, multilingual pretrained representations and retrieval of external factual evidence as possible future directions. Within the reported evaluation, however, SCSNet’s central result is that weighting early replies by estimated contribution and using gated, stage-wise state updates can improve rumor detection before a propagation network is complete, reduce reliance on event-specific patterns and support transfer to unseen events.

Subject of Research: Cross-event early rumor detection by weighting sparse social media replies and updating event states in stages

Article Title: Early detection of online public opinion rumors for cross-event generalization

Article References: Kang, K., & Yuan, Z. (2026). Early detection of online public opinion rumors for cross-event generalization. Discover Artificial Intelligence, 6(1), Article 1173. https://doi.org/10.1007/s44163-026-02227-7

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02227-7

Keywords: early rumor detection, cross-event generalization, SCSNet, social media replies, propagation modeling, PHEME, Twitter15, Twitter16, Twitter-COVID19, neural networks, Early, detection