[2015年第2期] 在线社交网络的可视化分析

visual analysis of online social network

        黄河燕

         社交网络，可视化建模，社会计算，主题分析，用户个性化

        social network， visualization， social compurting， topic analysis， user personalization

        随着在线社交网络规模的指数性增长， 可视化分析成为理解社交媒体 及用户的重要手段。 文章针对社交媒体、社交用户 、社交媒体与用户的交互 3 个 角度， 从社交媒体主题可视化建模、社交用户偏好可视化建模、社交网络个性化 检索可视化 3 个方面，分析社交网络可视化技术的研究成果。 提出时空融合可视 化、“三网 ”融合可视化等未来国内研究方向。 以期利用在线社交网络， 为互联网 服务、国家安全提供有力支撑。

        With the exponential growth of online social networks， the visual analysis technique has become an important research issue for understanding social media and social network users. In this article, targeting at social media, social network users, and their interactions respectively, we provide a detailed survey on research work of social media topic analysis, social user modeling, and personalized social information retrieval. We hope these techniques can be utilized to improve the Web user experience and the national security service. Recent research work in visualizing social media are divided into：（1）topic analysis;（2） topic diversification; and（3） topic evolution. Topic analysis means to construct a group of themes from a set of documents, and assign each document with a theme or a group of themes. Typical methods include the Probabilistic Latent Semantic Indexing（PLSI）, the Latent Dirichlet Allocation（LDA）, etc. Topic diversification refers to distinguish different types of topics. For example, some topics are objective contents, and other topics are subjective sentiments. In order to achieve this, supervised topic models are proposed to identify different topic types. Topic evolution is to connect topics in consecutive time intervals through topic similarity measures, through which one can see the topic change, the topic merging, and the topic separation. Generally, KL-divergence is utilized as the similarity measure. The work in visualizing social network users are classified into：（1） single-domain user preference modeling;（2） cross-domain preference transformation; and（3） user preference evolution. Single-domain user preference modeling is to describe a user’ s personal bias on items in a single domain. For example, to model a user’ s preference in a movie review system. Typical solutions are to combine the review aspect modeling with traditional collaborative filtering, in which the aspect is utilized for preference description. Cross-domain preference transformation is to learn the user preference pattern through multiple domains. In a typical scenario, a user has some feedbacks in a movie review site, but has very few feedbacks in a book site. By transferring learning techniques, the movie feedback is utilized to inference the book preference. User preference evolution refers to distinguishing a user’ s interests in different time periods. For example, a user may pay attention to furniture when he/she moves to a new house, but he/she may be interested in milk powder when he/she has a baby. By modeling the evolution patterns, the interests in the next time period can be predicted. The work in visualizing the interactions between social media and users are grouped into：（1）personalized search;（2） personalized recommendation; and（3）search/recommendation results reconstruction. Personalized search means to return search results by considering a user’ s search history. Thus a user model is constructed in advance, and the displayed search results are required to match the user’ s intent. Personalized recommendation is to retrieve information without queries. Usually, collaborative filtering models are utilized to match a user’ s interest and the relevant items. In visualizing the items to the user, the diversity among a group of displayed items needs to be considered. Therefore, some entropy-based objectives are proposed to be joint with the traditional collaborative filtering objectives. Search/recommendation results reconstruction refers to convert unstructured retrieval results into structured data automatically. This will make both conveniences for user experiences and search engine indexing. Some future work directions are also discussed in the end, such as visualizing temporal-spatial properties of social network, visualizing the combination of internet of things, mobile internet, traditional internet, etc.