基于多主体联合学习的社交网络谣言鲁棒辨识研究

在线社交网络是一种复杂的非线性、有限置信空间，多模态特征组成了异质的多域信息网络，各主体的活动模式具有动态随机性，从而难以掌握谣言的传播发酵规律以及实现高效的辨识。针对舆情传播网络的特征复杂性及情境时变性问题，本项目基于多主体联合学习，研究社交网络谣言辨识模型及其鲁棒优化方法。首先，通过分析舆情传播网络中多域特征的协同作用机制，研究基于图学习的交叉映射模型，实现深度表示与异质融合；然后，揭示多主体活动的有限置信性以及不确定性机理，研究基于多主体联合学习的谣言辨识模型，借助交互协同训练提升辨识效率；最后，引入实时数据流形成动态反馈的在线学习场景，研究基于强化学习的模型参数自监督优化方法，构建起具有较强泛化性能的鲁棒辨识架构。本项目旨在解决在线社交网络中的谣言辨识问题，研究成果能够丰富该问题的理论研究体系，在网络空间的舆情治理方面具有应用价值。

Online social networks are a kind of intricately nonlinear space as well as finitely confidential space. On the inside, multimodal features constitute heterogeneous multi-field information network, and the activity patterns of each agent are dynamically random. Hence, it is difficult to grasp the propagation and fermentation law of rumors, so that efficient identification cannot be realized. The problem faced here lies in feature complexity and contextual temporality of the opinion communication network. On the basis of multi-agent federated learning, this project investigates an identification model for cyber rumors and the robust optimization method. Firstly, by analyzing the synergy mechanism of multi-field features in opinion communication networks, the graph learning-based cross-mapping model is studied to achieve deep representation and heterogeneous fusion. Then, the finite confidence and uncertainty mechanisms of multi-agent activities are revealed, and the rumor identification model based on multi-agent joint learning is studied. Thus, identification efficiency can be improved with the assistance of interactively collaborative training scheme. Finally, real-time data streams are introduced to online learning scenarios with dynamic feedback. The self-supervised optimization method of model parameters based on reinforcement learning is studied, so that the robust identification architecture with strong generalization performance is constructed. This project aims to solve the problem of rumor identification in online social networks. The research results are able to enrich the theoretical research framework of this problem, and have applicable value in the governance of public opinion in cyberspace.