Early detection of online public opinion rumors for cross-event generalization
TL;DR
Results indicate that the proposed SCSNet method is more reliable for unseen-event generalization and sparse early-propagation judgment, and shows consistent performance improvement from 10 to 30 min across repeated runs.
Abstract
Early detection of online public opinion rumors aims to identify potential risks before propagation has fully unfolded. However, existing methods often rely on relatively complete propagation structures, stable event-specific vocabulary, or sufficient reply contexts, and thus tend to produce unstable judgments when early information is sparse and event distributions vary substantially. To address this problem, we propose the Stage Contribution State Network (SCSNet). The method first estimates the effective contribution of early replies to the current judgment from reply content, arrival time, and propagation depth, and then organizes replies with similar contribution levels and information changes into stage representations. In this way, it highlights key cues such as questioning, supplementation, and clarification, while reducing the interference of repeated reposts, emotional expressions, and irrelevant comments. On this basis, SCSNet progressively updates the event state through a linear state update mechanism and outputs early predictions at any observation window. Experimental results show that SCSNet achieves consistent improvements on Twitter15, Twitter16, and PHEME. In particular, it achieves a Macro-F1 of 82.74% in the cross-event test on PHEME, improving over the closest competing method by 1.24%, and shows consistent performance improvement from 10 to 30 min across repeated runs. These results indicate that the proposed method is more reliable for unseen-event generalization and sparse early-propagation judgment.