Media Popularity Prediction Algorithm Based On Multiple Attributions

With the continuous digital transformation of traditional media in recent years,digital media content has gradually become the principal part of the Internet.However,only a small part of it is extremely popular,so timely and effective popularity prediction methods are needed to identify the ones that will be popular.Popularity prediction researches are valuable in many applications,such as advertising,network management,cache strategy and so on.There are many factors affecting popularity,and multiple attributions of digital media content are the most important ones when we lack historical data,therefore we pursue our research on media popularity prediction algorithm based on multiple attributions,and we take posts and videos as main objects of our research.Firstly,we focus on the research of relative popularity prediction for posts.Popularity can be strongly dependent on many non-content factors,but content-factors should not be dismissed.Complicated interplay of multiple modes in posts makes it difficult to predict relative popularity.In this paper,we propose a multi-modality model to predict relative popularity.Our model uses deep neural network and attention mechanism to extract image feature and text feature from a post.Then multi-task learning is introduced into our model,on the one hand,our model retains relevance between image and text in the same post,on the other hand,our model implicitly constructs interplay between image and text,and predicts relative popularity through a pair of posts.Experiments on real data show the effectiveness of our proposed models.Then we focus on the research of initial popularity prediction for user-generated videos(UGVs).Previous work has demonstrated that UGVs generally receive the highest attentions during the first few days and initial popularity is highly correlated to future popularity.We take data from a Chinese online video service provider named after Youku as object for the first time,and use multiple attributions of metadata to construct video property features,user property features and textual analysis features as predictors.In the meanwhile,we use four basic regression models to build our initial popularity prediction model,these basic models are SVR-L,SVR-RBF,neural network and XGBoost.Results with two index prove the effectiveness of above three sets of features and demonstrate that XGBoost has the highest performance.Through feature importance analysis,we find that one important feature in forecasting initial popularity of Youku videos is category,which does not contribute sufficiently in YouTube videos,in addition,we find that sentiment features have little significance.