Structure-Aware Propagation Graph Learning for Platform-Side Rumor Risk Screening Under Limited Observation

Online platforms need scalable risk-screening methods for rapidly spreading rumors, misleading content, and large-scale user responses. Platform-side content governance often needs to rank suspicious events when only partial propagation evidence is visible, so that limited review resources can be assigned first to high-risk events. This article focuses on that setting. We propose a Structure-Aware Bidirectional Graph Convolutional Network (SA-BiGCN), which combines a BiGCN-style propagation graph encoder with structural statistics computed from the currently visible propagation tree to estimate event-level rumor probability. We construct Main Weibo V2 from 4664 raw Weibo source posts and their propagation structures, encode node text with a unified vocabulary, and use fixed event-level splits. The model evaluates source-post veracity at Top 10, Top 30, Top 50, Top 100, and full observation windows according to a timestamp-audited propagation order, using only the currently visible propagation-tree subgraph. The experimental results show that SA-BiGCN reaches 92.72% Avg F1 and 91.28% Worst F1 on Main Weibo V2 while using one shared checkpoint for all observation windows. Feature-set, leave-one-feature-group, sampling-strategy, topology-corruption, deployment-cost, and transfer analyses indicate that structural statistics computed from the currently visible propagation tree provide auxiliary evidence mainly in early-window and weakest-window settings. We further extend SA-BiGCN with graph-attention, temporal-quality, response-style, window-aware, and RootRoBERTa fusion variants. These additional cues are retained as diagnostic trends and do not replace the main model. Overall, SA-BiGCN is best understood as a lightweight single-checkpoint propagation-tree scorer.