Analysis and Prediction of Content Popularity for Online Video Service:A Youku Case Study

Understanding the characteristics and predicting the popularity of the newly published online videos can provide direct implications in various contexts such as service design, advertisement planning, network management and etc. In this paper, we collect a real-world large-scale dataset from a leading online video service provider in China, namely Youku. We first analyze the dynamics of content publication and content popularity for the online video service. Then, we propose a rich set of features and exploit various effective classification methods to estimate the future popularity level of an individual video in various scenarios. We show that the future popularity level of a video can be predicted even before the video's release, and by introducing the historical popularity information the prediction performance can be improved dramatically. In addition, we investigate the importance of each feature group and each feature in the popularity prediction, and further reveal the factors that may impact the video popularity. We also discuss how the early monitoring period influences the popularity level prediction. Our work provides an insight into the popularity of the newly published online videos, and demonstrates promising practical applications for content publishers,service providers, online advisers and network operators.

Analyzing the dynamics of online video popularity

Given the large volume of video content and the diversity of user attention, it is of great importance to understand the characteristics of online video popularity for technological, economic and social reasons. In this paper, based on the data collected from a leading online video service provider in China, namely Youku, the dynamics of online video popularity are analyzed in-depth from four key aspects: overall popularity distribution, individual popularity distribution, popularity evolution pattern and early-future popularity relationship. How the popularity of a set of newly upload videos distributes throughout the observation period is first studied. Then the notion of active days is proposed, and the per-day and per-hour popularity distributions of individual videos are carefully studied. Next, how the popularity of an individual video evolves over time is investigated. The evolution patterns are further defined according to the number and temporal locations of popularity bursts, in order to describe the popularity growth trend. At last, the linear relationship between early video popularity and future video popularity are examined on a log-log scale. The relationship is found to be largely impacted by the popularity evolution patterns. Therefore, the specialized models are proposed to describe the correlation according to the popularity evolution patterns. Experiment results show that specialized models can better fit the correlation than a general model. Above all, the analysis results in our work can provide direct help in practical for the interested parties of online video service such as service providers, online advisers, and network operators.

基于深度信念网络的在线视频热度预测

针对在线视频热度预测研究中分类及预测效果欠佳,规则化较多和较缺乏实践检验等问题,通过对实际在线视频服务系统所采集的海量数据研究,提出一种基于深度信念网络(Deep Belief Networks,DBNs)的视频热度预测方法。首先,结合社交网络的关注度和视频关键词的搜索热度,对影响因子进行了建模和量化处理;其次,根据输入和输出变量确定了DBNs各层网络的结构,优化了网络参数和预测模型;最后,通过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方法在视频热度预测上准确率最高79.47%(国内视频)、65.33%(国外视频),可以为在线视频上映前的投资、宣传以及风险评估提供较全面可靠的参考决策。