基于异质图注意力网络与多特征融合的跨社交媒体用户识别研究

User Identification Across Social Media Based on Heterogeneous Graph Attention Network and Multi-Feature Fusion

跨社交媒体用户识别对于网络舆情的协同治理以及用户偏好的全方位识别与预测具有重要的指导意义。针对现有方法存在数据表达能力弱、忽略用户信息的动态性和关联性的问题,本文提出一种融合异质特征嵌入与实体动态关联的跨社交媒体用户识别模型。首先,整合用户的基本属性、生成内容和社交结构信息,构建各个社交媒体的异质信息网络;其次,通过设计新的元路径识别策略构造邻接矩阵,使用异质图注意力网络模型汇聚用户节点信息,增强节点特征的表示能力;再其次,引入了3种连续时间衰减函数,对跨社交媒体的实体相似矩阵进行加权,增强实体之间的动态关联;最后,融合单社交网络和跨社交网络中的以上特征,利用多层感知机实现跨社交媒体用户的识别和预测。在微博-知乎真实数据集中的研究结果显示,本文模型的整体性能优于其他基准模型,特别是线性衰减函数,其展现了最佳效果,且本文提出的元路径识别策略对提升用户识别效果具有重要作用。

Cross-social media user identification is crucial for guiding the collaborative governance of online public opinion and for comprehensively identifying and predicting user preferences.This study introduces a model for recognizing cross-social media users that integrates heterogeneous feature embedding and dynamic entity association to address the issues of weak data representability and neglect of the dynamic and associative nature of user information in current methods.The heterogeneous information networks of various social media platforms were created by incorporating basic user attributes,content generation,and social structure information.A new meta-path recognition strategy was designed to construct an adjacency matrix,allowing the heterogeneous graph attention network model to aggregate user node information and enhance the representability of node features.Additionally,three continuous time decay functions were introduced to weigh the entity similarity matrix across social media platforms,enhancing the dynamic relationship between entities.By integrating features from both single and cross-social networks,a multi-layer perceptron was utilized to achieve the recognition and prediction of cross-social media users.Experiments conducted on the real Weibo-Zhihu dataset showed that the overall performance of the model was superior to that of other benchmark models.The linear attenuation function was found to have the most significant impact,and the meta-path detection strategy proposed in this article played a pivotal role in improving detection effectiveness.