Research On Popularity Prediction Of Micro-videos

With the rapid development of information technology,the form of media data we accept has changed from a single text data to a variety of multi-modal data with more vivid forms and richer contents.At the same time,with the popularization of various digital information acquisition devices and the Internet,micro-video,which is a data form taken,produced and shared by users,has become the newest and most popular one.For both users and platforms,how to predict the popularity of micro-videos has important practical significance in advertising,video recommendation and bandwith allocation.Based on these factors,this paper studies the problem of popularity prediction of micro-videos.Through research on the related technology of single-modal and multimodal media data analysis,we extract features of the micro-videos from mutiple modalities and propose two models for the popularity prediction problem.The main work is summarized as follows:(1)For the micro-video samples we used in this paper,we extract their features from multiple modalities and introduce the reasons,characteristics and extraction methods of each feature.Finally,for each short video sample,features from four modalities are extracted including visual features,textual features,acoustic features and social features.(2)A popularity prediction model based on low-rank representation of the feature matrix and sparse regression process is studied.First,we perform the low-rank selfrepresentation of the feature matrix to explore the low-rank structure information.Then,the two optimization problems are jointly optimized by combining the sparse regression process with the popularity score prediction.Finally we validate the effect of different features in the process of popularity prediction through experimental results.(3)Considering the particularity of social features in popularity prediction,a feature-discrimination transductive model is studied.The process of popularity prediction is divided into two stages: classification of popularity levels based on social features and popularity prediction based on the fused multi-modal features.Finally,by the measurement of normalized mean square error and spearman rank correlation,we prove the validity of the popularity level classification process in the model.