Research On Video Popularity Prediction Algorithm Towards Online Video Services

Today's online video services are facing the challenges from massive requests and huge scale of service data.In order to alleviate the service pressure,online video services providers usually develop various service strategies to improve the efficiency and economic benefits of the services.To ensure the effectiveness of service strategies,it is critical to predict videos' future popularity.The task of video popularity prediction is to infer the views of videos in future based on the accessible information obtained in the early stage of their lifetime.Currently,there are three core issues in the study of video popularity prediction: the first one is how to estimate videos' future views based on the the early data of the video;the second issue is the real-time online popularity prediction which requires the stability of the prediction performance with the time consumption of prediction controlled;the last issue is the popularity prediction of fine-grained video content whose popularity can hardly be predicted with methods for entire videos.In this dissertation,we explore to solve the above issues to improve the performance of video popularity prediction and the practicability of the technology.Our first work focuses on video popularity prediction solely based on early views.Early models on video popularity prediction prefer to construct the mapping between the early and future accumulative views.However,as the early accumulative views fail to indicate the future trend of views,these models are usually of low performance.To address this problem,we extend the accumulative views of videos into series against the age of videos to obtain more detail dynamics of views.By connecting the early and future dynamics of views,we are able to infer the future trend of views effectively.In particular,as the seriess of accumulative views are monotonic non-decrease,we model them with Pure Birth Process(PBP).The proposed VCDM model then infers the future views of videos by estimating the growth of videos' accumulative views in future based on their recent growth.This work is not only the start of our study,but also one of the earliest studies on timeseries based video popularity prediction.Experimental results show that the VCDM model outperforms the earlier proposed views-mapping based prediction models.With the deepening of social influence to the viewing behaviors of users,un-popular videos are more likely to have similar early views as popular ones which reduces the effectiveness of early views based popularity prediction.To address this problem,existing studies usually incorporate massive multi-modal features to help the prediction of videos' future views.However,most of these models fail to improve their performance on the prediction of popular videos while their complexity keeps increasing,largely reducing the practicability of models.We discover that the main reason for such phenomenon is that most features are weakly correlated to videos' views.Meanwhile,most existing studies ignore the fusion of features and have features discrete in models which limits the contribution of features to the prediction.Therefore,in the second work of our study,we first choose two groups of information which are closely related to videos' views for feature extraction: one is the data that can clearly reflect the attitude of users towards videos,the other is the sharing times of related social content of videos.Through the analysis on related data,we discover that the relative growth of videos' views is approximately Rayleigh distributed against the proportion of earlier viewers with positive and negative attitude towards the videos.Based on the discovery,we construct the early rating patterns with better discriminative power on the future trend of views.On the other hand,we learn the social features from the sharing data of videos' related social content through transfer learning.Finally,we introduce a video age-sensitive balancing mechanism to maximize the contribution of early rating patterns and social features in prediction.In the experiments,the proposed MFDI model performs better than the state-of-the-art existing models,especially on the prediction of popular videos.As a topic sourced from online video services,real-time online prediction is one of the core issues in related researches.Real-time means that the prediction should be completed in given time while online prediction requires models to maintain stable performance when updating the prediction results.Our previous works can help dealing with the limitation of accessible data,hence how to stabilize the online prediction performance is the task of our third work.Existing models usually complete the job by updating themselves with videos' views in the target period of recent prediction.However,for promptly service,it is hard to obtain the required views for model update which makes the model fall behind the views data.To address the problem,we construct a novel online prediction model IEVP with a recent discovered feature that is closely related to videos' views: the average watched percentage of video content.Follow the idea of our previous work,by discovering the relation between average watched percentage and video views,we estimate the relative growth of views in future based on the average watched percentage over a recent period.The prediction is achieved by combining the estimation of relative views growth and the views that occurs in the same period as the average watched percentage.Specially,as the temporal range of average watched percentage applied for relative views growth estimation is much shorter than that of views prediction,the estimation can be timely updated,thereby stabilize the online prediction performance of the whole model.The effectiveness of the proposed IEVP model is proven by the reported experimental results on real-world service data of i QIYI.Our last work is to solve the problem of video segment popularity prediction which is a brand new task coming with the refined media content service.In online video services,video segments are the pieces of video content divided with certain duration.The biggest challenge of video segment popularity prediction is that besides videos' first segments,the rest segments are always skipped by users during the viewing process.Thus,for most segments,their early views can no longer reflect the future views effectively.Meanwhile,although the views of video segments are closely related to the viewing traces of single users,it is infeasible to investigate the viewing traces of all the users in real-time services.Towards this end,we distinguish the popularity prediction of a video's first and rest segments and design two specific prediction models.In particular,we treat the popularity prediction of videos' first segments as the same task for entire videos and fulfill the job with a widely-adopted Multi-Linear Regression Model.On the other hand,a statistic-leveled non-linear prediction model is designed for the rest segments based on the popularity of recent viewed previous segments.Moreover,an online compensator is introduced to equip the two prediction models with online learning capability,stabilizing the online prediction performance of the two models.The proposed method has been deployed in i QIYI Video CDN(VCDN) and is now benefiting the service efficiency.