| Video on demand(VoD)service is popular,and the competition among video service providers becomes more and more fierce.Aside from improving users' experience,and attracting more potential users,video service providers are also obliged to optimize the allocation of resource.In order to achieve these goals,it is essential for video service providers to understand the users' behaviors in their lifespan,as well as the evolution of video popularity.However,a systematical study on these problems in VoD system is desired but absent in literature.In this paper,based on a large-scale dataset of user watching behavior from PPTV,machine learning methods are applied to predict user lifespan based on user first-week behavior records.In addition,we observe video popularity dynamic characteristics and predict video popularity evolution patterns based on its initial daily view counts and meta-data of videos.The contributions of this paper are as follows.(1)We analyze user watching behavior in their lifespans based on a large-scale dataset from PPTV.We observe that user's behavior features such as visit frequency,number of views and watching finishing ratio basically show inverted U-shaped trajectories during user lifespans.A new preference feature,i.e.,Popular Topic Preference(PTP),is introduced to represent a user's preference for popular video.We observe that a user's PTP decreases with time.The longer a user stays in the system,the less popular videos she watches.(2)Machine learning methods are applied to predict whether a user will have a long lifespan in VoD system based on the video watching behavior records in her first week in the system.Experimental results show that the PTP feature helps to improve the F1-score of prediction by 8.8%,and reach 0.74 at the best,which is relatively high.Moreover,the most relevant feature is the visit frequency,and PTP is more important than finishing ratio.(3)We explore how a video's popularity evolves over time and find that the evolution process of popularity falls into 'burst' and 'non-burst' patterns.Videos of the'burst' pattern may keep popular for no more than a couple of weeks while those of the'non-burst' pattern are active for a long time.To classify videos into 'burst' and 'non-burst' pattern in terms of popularity evolution,an effective metric named popular-span is proposed to measure the duration of the popular period of a video.A method is also proposed to measure the popular-span instead of the lifespan by considering peaks and valleys of videos' popularity curves.(4)We predict a video's popularity evolution pattern based on its initial daily view counts and meta-data of videos,including user rating,number of raters and video genres.The experimental results show that meta-data information helps to improve the F1-score of popularity evolution pattern prediction by 9%,and reach 0.85 at the best.A 'Popularity Evolution Pattern Aware Caching'(PEPAC)scheme is proposed as an example of the application of popularity evolution pattern prediction.Experiments results show that under the precondition of improving the cache hit ratio,the cache replacements can be reduced by 4.8%at most via PEPAC. |
PDF Full Text Request