Optimizing Video-on-Demand Systems Based On User Behavior Analysis

With the fast development of network technology including both hardware and software, the main content delivered on Internet has been changed from text to mul-timedia. The video-on-demand (VoD) service who contributes the majority of traffic has become the most popular application on Internet. In the past few years, lots of re-search works and commercial systems emerge in large numbers. It is challenging to support the VCR functionality, especially the freewill jumps, while maintaining a s-mooth streaming quality in VoD systems within the current network conditions. How to minimize users’ wait time while guarantee a certain playback fluency for improving user experience urgently need to be addressed.A customer in a VoD system is not just a passive user but a participator, he can produce VCR operations to affect the system performance. In this paper, we analyze users’ viewing behaviors to optimize the VoD system designs for improving users’ex-periences. Specifically, the contributes of this paper are summarized as follows:1. A method of mining users’ interests inside a video is proposed. Through analyz-ing the statistics of log files of a real-world VoD system, we find that users may produce seeking operators with two different underlying meanings and they can’t be treated in the same way. So we classify users’ jumping behaviors into two kinds:satisfactory and unsatisfactory jumps. Based on the two classification probability, we score every video segment which can represent the levels of users’ interests in different positions of the video. By applying to different kinds of videos, the results show our scoring algorithm can capture the users’ interests accurately.2. We present a bookmarking algorithm for automated marking the highlights of a video. Based on users’ interests, the algorithm can automatically bookmarking different kinds of videos’ highlights which can be utilized to guide users’ seeking and improve experiences. Applying the algorithm on a large number of videos, the mark results prove this algorithm can get better performance than the traditional video semantic analysis methods which are based on image processing, including accuracy, universality, stabili-ty, efficiency and etc. We also give an application example of highlight bookmarks:fast preview of video hotspots in BitTorrent downloading. The simulation and PlanetLab experiments prove the feasibility of our strategy.3. A dynamic start-up threshold algorithm of playback is addressed. As users’ viewing time distributes extremely uneven, the existing fixed start-up threshold method can’t adapt to users’ viewing behaviors. Based on this observation, we propose a dy-namic start-up threshold algorithm. On the basis of the Gaussian traffic, our algorithm combines users’ viewing interests with current download bandwidth for dynamically calculating the optimal threshold which can minimize users’ wait time while ensure a specified playback fluency. If we use this threshold to control users’buffer length, the traffic consumptions of clients can be reduced which is very important in mobile networks. The simulation experiment using real log files proves the effectiveness of dynamic start-up threshold.