现代电子技术2025，Vol.48Issue(24)：47-53,7.DOI:10.16652/j.issn.1004-373x.2025.24.008
融合边权自适应与层级专家混合的谣言检测模型ERLMGcn
Edge-reweighted layer-wise mixture-of-experts graph convolutional network ERLMGcn
摘要
Abstract
In allusion to the challenges in rumor detection models for social networks,namely noisy edge relations,the lack of multi-scale feature aggregation,and the over-smoothing of node representations,an edge-reweighted layer-wise mixture-of-experts graph convolutional network(ERLMGcn)is proposed.Text representations are first extracted by means of bidirectional encoder representations from transformers(BERT),while metadata is encoded with general strategies.The two are then concatenated and linearly projected into unified node features.In each graph convolutional network(GCN)layer,an adaptive edge-weighting mechanism is introduced to highlight critical propagation links and suppress noise during message passing.A gating network is employed to selectively aggregate node representations along the layer dimension,thereby capturing both local and global features while alleviating over-smoothing.The comparative and ablation experiments were conducted on the Ma_Weibo and CED_Dataset.In comparison with the best-performing graph-based baseline,the proposed model can improve accuracy by 8.43%and 3.39%on the two datasets,respectively,and also achieve consistent gains across other metrics.The ablation results further verify the effectiveness of the adaptive edge weighting and the layer-wise mixture-of-experts gating mechanism.It provides an effective solution for rumor detection in Chinese social media.关键词
谣言检测/图神经网络/新浪微博/专家混合/层级融合/图卷积网络/自适应边属性/谣言识别Key words
rumor detection/graph neural network/Sina Weibo/mixture-of-experts/layer-wise fusion/graph convolutional network/adaptive edge attribute/rumor identification分类
信息技术与安全科学引用本文复制引用
潘杰,王娟,王楠..融合边权自适应与层级专家混合的谣言检测模型ERLMGcn[J].现代电子技术,2025,48(24):47-53,7.基金项目
河北省社会科学基金项目:大数据驱动的京津冀社会安全风险计算与智能决策(HB22SH011) （HB22SH011）
