面向在线视频服务的播放量预测算法研究

发布时间：2019-05-10 19:07

【摘要】：随着网络视频爆发式增长,在线视频服务资源面临着严重过载,准确预测视频播放量对供应商而言越来越重要。论文通过对实际在线视频服务系统所采集的海量数据研究,针对视频放映的不同时间段,分为上映前的精准预测和上映后的同步预测二个阶段:1)视频上映前,针对传统预测模型分类和预测效果不佳、规则化较多和缺乏实际应用价值等问题,提出一种基于深度信念网络(Deep Belief Networks,DBNs)的视频播放量预测方法。首先,结合社交网络的关注度和视频关键词的搜索热度,对影响因子进行建模和量化处理;其次,根据输入和输出变量确定DBNs各层网络的结构,优化网络参数和预测模型;最后,利用在线视频服务商的数据对深度信念网络进行训练,并多次交叉实验对比分析,结果表明基于DBNs方法在视频播放量预测准确率上有较大提升,有效实现了视频播放量的早期预测。2)视频上映后,通过对在线视频早期播放量时序的统计分析,提出一种基于ARMA模型的视频播放量预测方法,同步预测视频未来某天的播放量。根据视频播放量时序特征的差异性选择不同的预测模型,模型在对非平稳的国内视频和季节性明显的国外视频日播放量的同步预测获得了较高精确度,相比传统的移动平均法、指数平滑法和最小二乘法的预测方法获得了明显的提升,具有实际的参考价值。通过对深度信念网络和时间序列模型的研究,本文实现了在不同时间阶段对视频播放量进行及时、持续、准确的预测,既能为视频上映前的投资、评估提供较全面可靠的参考决策;又能够得到上映后未来时间点精确的播放量波动范围,为设计合理的广告投放、资源存储和商业决策提供支持。
[Abstract]:With the explosive growth of online video, online video service resources are facing serious overload, so accurate prediction of video playback is becoming more and more important for suppliers. Based on the research of massive data collected by the actual online video service system, this paper is divided into two stages: accurate prediction before release and synchronous prediction after release according to the different time periods of video screening: 1) before the release of video, In order to solve the problems of poor classification and prediction effect of traditional prediction model, more regularity and lack of practical application value, a video playback prediction method based on deep belief network (Deep Belief Networks,DBNs) is proposed. Firstly, the influence factors are modeled and quantified according to the attention of social network and the search heat of video keywords. Secondly, the structure of each layer of DBNs network is determined according to the input and output variables, and the network parameters and prediction model are optimized. Finally, the data of online video service providers are used to train the deep belief network, and many cross experiments are compared and analyzed. The results show that the accuracy of video playback prediction based on DBNs method is greatly improved. The early prediction of video playback is effectively realized. 2) after video release, through the statistical analysis of the timing of online video early playback, a video playback prediction method based on ARMA model is proposed. Synchronously predict the amount of video to be played one day in the future. According to the difference of time series characteristics of video playback, different prediction models are selected. The model obtains higher accuracy in the synchronous prediction of non-stationary domestic video and seasonally obvious foreign video daily broadcast volume, compared with the traditional moving average method. The prediction methods of exponential smoothing method and least square method have been improved obviously and have practical reference value. Through the research of deep belief network and time series model, this paper realizes the timely, continuous and accurate prediction of video playback at different time stages, which can not only invest in video before release. The evaluation provides a more comprehensive and reliable reference decision; It can also get the accurate fluctuation range of broadcast volume at the future time point after release, and provide support for the design of reasonable advertising, resource storage and business decisions.
【学位授予单位】：深圳大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41

【相似文献】