基于二分图的个性化图像标签推荐算法  

A personalized image tag recommendation algorithm based on bipartite graph model

作　　者：赵天龙 刘峥[1,2] 韩慧健 张彩明[3,4] Zhao Tianlong;Liu Zheng;Han Huijian;Zhang Caiming(School of Computer Science and Technology,Shandong University of Finance and Economics, Ji'nan,250014,China;Shandong Provincial Key Laboratory of Digital Media Technology,Ji'nan,250014,China;Schoo 1of Software,Shandong University,Ji'nan,250101,China;Shandong Co-Innovation Center of Future Intelligent Computing,Yantai,264025,China)

机构地区：[1]山东财经大学计算机科学与技术学院,济南250014 [2]山东省数字媒体技术重点实验室,济南250014 [3]山东大学软件学院,济南250101 [4]山东省高等学校协同创新中心:未来智能计算,烟台264025

出　　处：《南京大学学报（自然科学版）》2018年第6期1193-1205,共13页Journal of Nanjing University（Natural Science）

基　　金：国家自然科学基金(61772309,61303090,61472221,61332015);国家自然科学基金浙江两化融合重点项目(U1609218);教育部人文社会科学研究项目(13YJC860023);济南市高校自主创新计划(201303012);济南市青年科技明星计划(201406001);山东省自然科学基金省属高校优秀青年人才联合基金(ZR2018JL022);山东省高校科研创新团队

摘　　要：传统的图像标签推荐方法通过对图像视觉内容的分析计算标签与图像的相关度,完成标签推荐任务.而社会网络图像具有丰富的元数据,例如图像所属群组、地理位置等,充分利用这些元数据对于提高标签推荐的准确性具有积极意义.提出一种基于二分图的个性化图像标签推荐算法,通过充分挖掘图像、群组、地理位置与标签的关系,针对用户提供的少量标签进行个性化图像标签推荐.该算法建立了图像-标签、群组-标签、地理位置-标签等三个二分图模型,考虑到每个标签的重要性不同,引入TF-IDF(Term Frequency-Inverse Document Frenquency)技术对标签进行加权处理.利用二分图将初始标签分值进行信息扩散,计算出最终标签分值向量,并将该向量中分值较高的标签作为推荐结果.实验结果表明,融合了图像与群组、地理位置等元数据的个性化图像标签推荐结果的NDCG(Normalized Discounted Cumulative Gain)值优于仅单方面考虑图像、群组以及地理位置的标签推荐结果.Traditional image tag recommendation methods mainly concentrated on the analysis of the visual content of the image and the calculation of the correlation between the tag and the image to achieve the tag recommendation task.However,social network images have rich metadata,such as the images’ group and site,etc.It is of great importance to make full use of these metadata to improve the accuracy of tag recommendation.In this paper,wepropose a personalized image tag recommendation algorithm based on the bipartite graph model.Furthermore,the proposed algorithm recommends personalized image tags for users with only a small number of user-supplied tags by fully exploiting the relationship between images,groups,sites and tags.In addition,the proposed algorithm constructs three bipartite graph models,such as image-tag model,group-tag model and site-tag model.As the importance of each tag is different,we introduce the TF-IDF(Term Frequency-Inverse Document Frenquency)technology to weight for each tag.The bipartite graph is used to spread the initial tag value,and the final tag value vector is constructed.Then,the tags with higher values in this vector are used as the recommendation results.Exploiting Normalized Discounted Cumulative Gain(NDCG)as performance evaluation criteria,experimental results demonstrate that the NDCG value of the proposed personalized image tag recommendation algorithm which combines three types of metadata information(such as image,group and site)is significantly higher than the NDCG value of the image tag recommendation results which only using image,group and site information individually.

关 键 词：图像元数据 标签偏好 二分图 个性化标签推荐 标签排序 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]