基于社交网络模型的视频分享关键技术研究

Video sharing service represented by YouTube has attracted widespread attention in recent years. However the sustainable development is severely hindered by the intrinsic limit of the C/S architecture. As an alternative promising solution, P2P streaming has been proposed for large-scale video sharing. Unfortunately, using P2P delivery for short video clips can be quite challenging. First, the P2P overlay will suffer from an extremely high churn rate, because the length of a YouTube video is short and a user often quickly loads another video when finishing the previous one. Second, the startup and playback delay would be effectively amplified from the perception of users for the shorter length. Third, there are a huge number of unpopular videos with few viewers, and thus many of the P2P overlays will be too small to function well. Moreover, P2P streaming would dramatically increase the traffic over inter-ISP links like other P2P applications. . In order to resolve above questions, our research will focus on the following aspects: (1) a locality-aware overlay construction mechanism based on social interest graph, which can efficiently alleviate the inter-ISP traffic and the dynamics of the overlay caused by switching the videos frequently. By long-time measuring the YouTube-like videos, we find that the videos have strong correlations with each other, so we can explore the interest correlation for enhancing the streaming quality. The new overlay is organized by the viewers with similar interest, so it is more stable and larger than the traditional scheme which is organized by the viewers sharing the same video; (2) a collaborative strategy based on auxiliary peers from other overlays, which can improve the viewing quality of unpopular videos and optimize the bandwidth allocation when the conditions is in overloaded; (3) a delay-aware scheduling mechanism to improve the playback quality and load-balancing; (4) a user preference guided prefetching and caching strategy that enables delay-minimized smooth transition between video playbacks; (5) performing extensive simulations and prototype experiments to evaluate the performance of the socialized video sharing system, and analyzing the system with the complex network model, based on the traces collected from the practical prototype system.

随着YouTube类网站获得极大成功，视频分享服务受到广泛关注。本项目针对当前P2P视频分享系统中存在的重叠网络高度动态性、资源分布不均匀、视频切换延时过长、流量频繁穿越ISP骨干链路等问题，将社交网络模型应用于视频分享系统中，拟在以下方面展开研究：（1）基于兴趣图谱和位置属性的重叠网络构建机制，以克服用户频繁切换视频导致的重叠网络高度动态性，缓解ISP域间流量压力；（2）基于知识量模型的跨重叠网络辅助节点协作策略，以优化系统的带宽资源配置，提高非热门视频用户的收看质量；（3）时效优先的数据调度策略，以保证短视频用户的收看质量、实现均衡负载；（4）基于用户偏好的视频预取与缓存机制，以改善用户的视频切换延时，扩大视频分享范围；（5）基于原型系统的特征测量数据，建立视频分享系统的复杂网络模型，研究其传播机理和运行规律，为视频分享系统的性能解释与评价模型提供理论证明。

随着YouTube类网站获得极大成功，基于社交网络的视频分享服务作为新型的互联网杀手级应用受到各国学者的广泛关注。本项目围绕目前视频分享系统中的关键技术展开研究，主要内容包括三个方面：（1）能够优化域间流量的P2P直播关键技术；（2）面向短视频分享的P2P点播关键技术；（3）基于行为的用户兴趣建模及视频收看概率预测算法。首先，项目组针对P2P直播系统提出了一个基于分层结构的重叠网络构建方法，用于建立视频服务器和用户节点之间、以及用户节点相互之间的数据传输连接。该方法通过控制系统用户建立的跨自治域的连接数量，达到优化域间流量的目的；通过限制数据包递交到每个用户节点的传输路径所经过的最大自治域数量，优化数据包递交延迟，从而保证直播视频的播放质量。此外，本项目还提出一个时效优先的调度策略，该策略在传统“拉数据”的基础上，制定了数据块请求的优先级规则，提出优先请求紧急数据块，以及滑动窗口中最新被允许请求的正常数据块。大量仿真实验表明，所提出的直播系统确实能够将ISP的域间流量控制在较低的水平，且保证用户高质量的视频收看效果。其次，项目组提出了基于兴趣图谱和缓存策略的短视频点播系统，小规模实验的结果表明通过引入视频缓存策略建立跨不同视频的稳定重叠网络，确实可以减小用户频繁切换短视频导致的高度动态性；此外通过增加每个用户的视频预取机制，针对下载带宽大的用户可以事先对感兴趣的视频进行预取，从而使得视频切换延迟大幅下降，目前缓存视频的命中率还较低，有较大的提升空间。最后，项目组单独对点播中的用户兴趣图谱和收看视频概率预测进行展开研究，提出了针对用户兴趣的关注关系提取及标注方法，以及基于图模型的用户兴趣图谱构建方法及收看视频概率的预测方法，方法简单有效，能够在一定程度上改善缓存视频的命中率。