Visual topic discovering, tracking and summarization from social media streams
Visual topic discovering, tracking and summarization from social media streams
复制标题
DOI：
10.1007/s11042-016-3877-1
复制
发表时间：
2017-04
影响因子：
3.6
通讯作者：
Zhao Lu;Y. Lin;Xiaoxian Huang;N. Xiong;Zhijun Fang
中科院分区：
文献类型：
--
作者：
Zhao Lu;Y. Lin;Xiaoxian Huang;N. Xiong;Zhijun Fang
如今，微博已变得流行，Facebook、Twitter、微博等社交媒体上的各种话题每分钟发布和分享的短信数以亿计。许多此类消息都包含捕捉人们生活中特定事件或时刻的视频。在这项工作中，我们寻求自动识别微博社交媒体流中发布的视频主题。虽然主题检测和跟踪（TDT）任务在多媒体检索中得到了广泛的研究，但由于内容短而嘈杂、主题多样且快速变化以及数据量大，自动发现、跟踪和总结社交媒体流中的视频主题仍然具有挑战性。在本文中，我们提出了一种基于 K 分图的方法来应对这些挑战。我们引入了 K 分图表示来同时对微博流中包含的视频、它们的纹理特征和视觉特征之间的关系进行建模。我们提出了一种新颖的联合聚类算法来捕获“关系聚类网络”（RCN）中 K 部分图的全局结构，其中将潜在元节点添加到网络中以表示视频聚类。基于该网络，我们提出了通过融合各种类型的特征和多种排名方案来跟踪和总结流中视频的方法。基于真实数据集的实验结果表明了我们的方法的有效性，与基线相比有了显着的改进。
Nowadays, microblogging has become popular, with hundreds of millions of short messages being posted and shared every minute on a variety of topics in social media such as Facebook, Twitter and Weibo. Many of such messages contain videos that captured particular events or moments in people’s life. In this work, we seek to automatically identify the video topics posted in the social media streams on Weibo. While Topic Detection and Tracking (TDT) task has been extensively studied in multimedia retrieval, automatically discovering, tracking and summarizing video topics from social media streams is still challenging due to short and noisy content, diverse and fast changing topics, and large data volume. In this paper, we propose a K-partite graph based approach to address these challenges. We introduce a K-partite graph representation to simultaneously model the relationships among videos contained in the Weibo streams, their textural features and visual features. We propose a novel joint clustering algorithm to capture global structure of the K-partite graph in a “relation cluster network” (RCN) where latent, meta-nodes are added to the network to represent video clusters. Based on this network we propose methods for tracking and summarizing the videos in streams through fusing various types of features and multiple ranking schemes. The experiment results based on a real dataset show the effectiveness of our method with significant improvement over baseline.