【作者】 张瑞；

【导师】 金志刚；

【作者基本信息】 天津大学 ， 信息与通信工程， 2018， 博士

【摘要】 为提高用户在社交网络中获取信息的效率和准确度,考虑了社交网络中的用户特征、内容特征、社区特征、异常特征等维度,提出了基于混合粒度的标签推荐模型,基于多维特征的话题发现与推荐机制,基于社区节点关系特征的动态社区检测策略,基于异常特征的话题倾向识别机制,从而解决了传统推荐系统研究中,用户兴趣多样性、内容主题提取、社区结构动态变化和话题倾向随时间演化对于话题主题发现与推荐结果的影响等问题。主要创新性研究成果如下:提出了基于混合粒度的标签推荐模型。针对已有推荐模型在社交网络用户标签运用中存在多样性、相关性较差的问题,将用户的可分析资源分解成由用户信息、标签和发表正文组成的混合粒度,在不同粒度上分别进行个人信息过滤及个性标签分析。从而计算用户标签的熵值与内联度,分类标注标签词汇,提取正文主题等,最终为用户推荐具有关联性较强的个性化标签。提出了基于多粒度的话题发现与推荐机制。针对社交网络中话题推荐准确度较低的问题,该机制对用户发表话题所包含的图像和文本,以及流行度等粒度以最大似然估计等方式进行计算,提高对于用户话题发现的准确度,以及对用户进行话题推荐的有效性。提出了基于社区节点关系特征的动态社区发现模型。针对社交网络中的内部节点频繁相互作用,复杂网络在不断动态演变的问题,通过快速搜索和密度峰值发现进行动态聚类,以提高真实复杂社区检测的准确性和适应性,从而通过社区结构关系来提高该社区内用户话题推荐的准确度。提出了基于异常特征的话题倾向识别机制。针对由于数据量的增加以及异常话题检测策略的更新,以及用户内容和行为为基础的传统异常话题倾向识别方法效果不断下降的问题,通过将文本进行情感分析后,结合质量控制的思想,通过检测可疑时间间隔和用户聚类分析,获得话题的异常倾向性。最后基于上述机制与策略,本文分析了现有社交网络所面临的挑战,设计了针对当前社交网络话题发现与推荐系统关键技术所提出的有效解决方案,并在真实数据集上验证了所提模型算法的可行性。更多还原

【Abstract】 In order to improve the efficiency and accuracy of getting information in social networks,this thesis considers the dimensions of user features,content features,community features,and abnormal features in social networks.It proposes a tag recommendation model based on hybrid grain,the topic discovery and recommendation mechanism based on multidimensional features,the dynamic community detection strategy based on the characteristics of community node relationships,and a spammer detection method based on sentiment analysis and quality control on topics.Thus,it solves the problems of traditional user recommended system research,the diversity of user interests,the extraction of content themes,the dynamic changes of community structure,and the tendency of topics to evolve over time.The main innovative research results are as follows:The existing recommendation model has a diversity and poor relevance in the use of social network user tags,thus a label recommendation model based on the hybrid grain is proposed.The user’s parsable resources are decomposed into a mixture of grain consisting of user information,tags,and published text.Personal information filtering and personality tag analysis are performed at different grain.Thus,the entropy and inline degree of the user’s tag are calculated,the tagging vocabulary of the tagging is categorized,the body theme is extracted,and finally the personalized tag with strong relevance is recommended for the user.In view of the low accuracy of topic recommendation in social networks,the thesis proposes topic discovery and recommendation mechanism based on multi-dimensions.The mechanism calculates the image and text included in the user’s posted topic,and popularity,in a manner such as maximum likelihood estimation,to improve the accuracy of the user’s topic discovery and the validity of the user’s topic recommendation.In view of the frequent interactions between internal nodes in social networks and the continuous dynamic evolution of complex networks,the thesis proposes a dynamic community detection model based on the characteristics of community nodes and performs dynamic clustering through rapid searching and density peak finding to improve the real complexity.The accuracy and adaptability of community testing to improve the accuracy of user topic recommendation through community structure relationshipsIn response to the problem of increasing the amount of data and the updating of anomaly detection strategies of topics,and the effect of traditional methods of identifying the tendency of abnormal topics based on user content and behaviors,a topic-based trend recognition mechanism based on anomalies is proposed.After the text was sentimentally analyzed,combined with the quality control,the abnormal tendency of the topic was obtained through the detection of suspicious time intervals and user cluster analysis.Finally,based on the above mechanisms and strategies,this thesis analyzes the challenges faced by existing social networks.An effective solution proposed for the key technologies of current social network topic discovery and recommendation systems is designed.The feasibility of the proposed model algorithm is verified on a real data set.更多还原