融合多网络社区身份的用户话题兴趣建模研究

With the rapid development of Web technology, online social media services have played a more and more important role in real life of users. Typically, a user has multiple social media accounts, i.e., Social Media Identity (SMI). Faced with the common phenomenon “one user, multiple SMIs”, how to develop effective cross-site topical interest models have become the priority for improving social media services and user experiences. Most of the existing studies focus on a single site, and it is not effective to model rich heterogeneous text information and user attributes from multiple sites, which might suffer from the data sparsity and bias. This project aims to deeply study how to model users’ topical interests based on the fact that a user is likely to have multiple SMIs. We propose three major candidate techniques to address this problem. First, in order to collectively link the SMIs of multiple users across sites, we develop a probabilistic factor graph model which can incorporate the information from a candidate linking pair itself and the correlations between candidate linking pairs. Second, by employing entities and entity relations as a bridge across sites, we build a probabilistic topic model for modeling users’ topical interests, which is flexible to characterize the variety of topical semantics across sites, the variety and consistency of the interests corresponding to multiple SMIs for the same users. Finally, based on the above built model with text information, we further incorporate the user attributes from multiple social media websites, which aims to improve over the basic model and enhance the explainability of user interest models. This project will be of great value to improve the user experiences of the social media sites which are featured on text-oriented service.

随着互联网技术的发展，社交媒体服务在真实生活中的作用日益重要，很多用户同时拥有多个网络社区帐号，即社区身份。针对“同一用户多社区身份”这一趋势，设计有效的跨社区话题兴趣模型成为改善社交媒体服务和用户体验的关键。已有工作主要面向单一社区的兴趣建模，不能充分融入多社区的文本语义信息与用户属性信息，可能具有数据稀疏性和有偏性等问题。本项目深入研究面向多社区身份的用户话题兴趣建模。通过同时刻画链指对自身的特征信息与链指对之间的关联关系，建立基于概率因子图的多帐号联合链指方法，更为精确地解决跨社区链指问题；通过引入实体与实体关系，设计面向多社区身份的话题兴趣主题模型，同时刻画话题跨社区表现形式的多样性、用户多社区身份的兴趣关联性与差异性；在文本信息的基础上，进一步在建立的主题模型中融入用户多社区的属性信息，改进兴趣建摸，加强兴趣归因解释。本项目将有力改善围绕话题兴趣打造的网络应用服务和用户体验。

随着互联网技术的发展，社交媒体服务在真实生活中的作用日益重要，很多用户同时拥有多个网络社区帐号。本项目深入研究了面向多社区账号的用户话题兴趣建模。首先，构建了跨社区链指数据集合和跨社区实体链指数据集合，为整个项目的顺利进行打下了数据基础。其次，研究用户兴趣所呈现的多种特征，包括用户兴趣数据的生活模式挖掘、用户兴趣数据的序列性挖掘和用户兴趣数据的话题摘要生成。第三，研究了融合多种背景信息的用户兴趣建模方法，主要包括基于多模态信息融合的用户兴趣建模、基于邻域信息的用户兴趣建模和基于受众信息的用户兴趣建模。第四，研究了跨社区用户兴趣建模算法，包括基于社交属性的跨社区用户兴趣建模、基于复杂特征映射的跨社区用户兴趣建模和基于异质信息网络表示的跨社区用户兴趣建模。第五，研究了用户兴趣的理解以及推荐系统可解释性，主要研究了基于知识信息的用户兴趣挖掘与建模。项目完成了原始制定的任务目标，为相关领域的研究和实践起到了一定的推动作用。