基于用户交互行为挖掘的个性化图像标签推荐研究

Achieving personalized search is the goal of the next-generation image search engine, of which the key problem is to give the image some tags that can not only meet user interests but also effectively describe image visual contents, that is, personalized image tag recommendation. Afterwards, a simple and efficient personalized image retrieval system can be implemented by taking advantage of the mature text retrieval technology. This project regards the image sharing community as application background, and considers user interaction behavior mining as the breakthrough point. Furthermore, the aim of this project is to study on an effective method of personalized image tag recommendation. Firstly, fuzzy synthetic evaluation is utilized to quantify the degree of user interaction and then user interaction graph is constructed. Secondly, the target user oriented community discovery which supports incremental expansion is implemented by considering the target user as constraints. Thirdly, all kinds of image metadata are fused by a modified topic model, and then the topic probability distribution of user interest and the time distribution of each topic are obtained. Hence, the user interest vector is constructed by a hierarchical vector space model based on the above two distributions. Finally, adopting image visual features and the interest vector of target user as priori conditions, the tags are recommended to the target user by selecting the tags which have maximum posterior probability from candidate tags using bayes theorem. Successfully implementing this project can promote the basic theory research of multimedia information retrieval and social network data mining, and provide core algorithms and key technologies for the next generation of image retrieval system.

实现个性化搜索是下一代图像搜索引擎的目标，其关键问题是为图像赋予既符合用户兴趣又能有效描述图像视觉内容的标签，即个性化图像标签推荐，进而利用成熟的文本检索技术实现简单高效的个性化图像检索。本项目以图像共享社区为应用背景，以用户之间的交互行为挖掘为切入点，研究个性化图像标签推荐方法。首先，通过模糊综合评价量化用户交互度，构造用户交互图；其次，以目标用户为约束条件，实现面向目标用户并支持增量扩展的社区发现算法；再次，通过改进的主题模型融合图像的各种元数据，求得用户兴趣的主题概率分布和各主题的时间分布，用层次向量空间模型将用户兴趣表示为"用户兴趣向量"；最后，以图像的视觉特征和目标用户的兴趣向量为先验条件，利用贝叶斯定理将候选标签库中具有最大后验概率的标签推荐给目标用户。本项目的实施，将对多媒体信息检索和社会网络数据挖掘的基础理论研究起到推动作用，为下一代图像检索系统提供核心算法与关键技术。

本项目旨在为图像赋予既符合用户兴趣又能有效描述图像视觉内容的标签（即个性化图像标签推荐），利用成熟的文本检索技术实现简单高效的个性化图像检索。本项目以图像共享社区为应用背景，以用户兴趣建模为切入点，研究个性化图像标签推荐方法；并开展了图像语义挖掘相关领域的研究工作，例如，图像处理以及自然语言处理。在本项目的资助下，项目组提出了一种面向图像共享社区的图像语义标注方法，并利用近邻投票算法计算图像与语义标签之间的相关度；提出了一种基于用户兴趣建模的个性化推荐算法；针对语义信息挖掘问题，提出了基于Apriori算法的KNN问句分类算法，提出了基于网页特征的邻位词共现算法，对候选关键词库进行扩展。.在本项目的资助下，项目组成员累计发表学术论文8篇，其中，SCI检索期刊论文1篇，EI检索期刊论文1篇，国际会议论文6篇。科研成果申请国家发明专利4项，培养硕士研究生7名。所取得的成果对于图像的语义信息挖掘，图像检索以及智能信息处理具有较高的应用推广价值，应用前景广阔。