基于异质信息网络表征学习的社交媒体事件异常检测

The Event Anomaly Detection from Social Media is the fundamental basis for the Public Opinion Monitoring and Guidance. In view of the massiveness, cross-modality, and complex attribute relationship issues on social media data, this project proposes a novel research strategy, which is "Event Detection - Representation Learning - Anomaly Detection", aiming at constructing event-related heterogeneous information network (HIN), and detecting event anomaly from social media based on the representation learning of HIN. Three key issues are addressed in this project. First, we construct event-related HIN in social media by establishing a consistent representation mechanism for multi-modal information and collaborating multi-modal information for event detection. Second, we learn the object and event representation based on the event-related HINs, combing with the network topology，contextual correlation and semantic relevance of objects within event-related HIN. Third, we detect the event anomaly and its key abnormal groups in social media on the basis of the representation of event-related HIN and the event correlation network. Above all, this project helps to enrich and improve the theories and methods regarding to Event Anomaly Detection from Social Media. What’s more, it contributes to overcome the problem of “Heterogeneous gap” in social media data, and provides theoretical and technical support for timely and accurate detection of social media event anomaly and prevention of misinformation spreading.

社交媒体事件异常检测是网络舆情监控与引导的重要依据。针对社交媒体数据海量、跨模态、属性关系复杂等特点，本项目提出了“事件发现-事件表征-异常判别”的研究思路，旨在通过构建事件相关异质信息网络，实现基于异质信息网络表征学习的社交媒体事件异常检测。拟解决的关键问题包括：一、建立同语义不同模态社交媒体数据一致性表征，实现多模态协同社交媒体事件发现，从而构建事件相关异质信息网络。二、基于异质信息网络表征学习，结合异质事件要素的上下文关联信息、语义相关性和网络拓扑结构，学习社交媒体事件及要素的表征。三、基于事件相关异质信息网络表征及事件关联网络，检测社交媒体异常事件，并进一步挖掘其中起关键作用的异常社团。本研究一方面有助于丰富和完善社交媒体事件异常检测研究的理论和方法；另一方面有助于克服社交媒体数据 “异构鸿沟”，为精准发现社交媒体事件异常、防止有害信息传播等提供理论与技术支撑。

社交媒体事件异常检测是网络舆情监控与引导的重要依据。针对社交媒体数据海量、跨模态、属性关系复杂等特点，本项目提出了“事件发现-事件表征-异常判别”的研究思路，旨在通过构建事件相关异质信息网络，实现基于异质信息网络表征学习的社交媒体事件异常检测。自2019年项目立项以来，项目组以社交媒体智能化治理为目标，探索基于异质信息网络表征学习的社交媒体异常检测方法，在跨模态一致性表征、异质信息网络表征学习和大规模图数据异常检测方面取得重要研究成果，并研制了面向高考网络舆情检测应用于历次教育部高考、研考等国家重大考试的信息安保工作。基于上述成果，在国内外重要学术期刊、会议上发表带项目61872287标注的论文共24篇，其中SCI检索13篇，CCF推荐国际期刊及会议论文8篇，申请和授权发明专利5项，培养博士生和硕士生8名。相关成果获得陕西省自然科学一等奖，排名第2/6，中国自动化学会科技进步特等奖，排名第14/21。