Short Video Popularity Prediction And Value Evaluation Based On Multi-modal Learning

With the rapid development of the Internet and New Media industries,short videos have completely changed the ecosystem of information dissemination with their concise and easy-to-understand nature,diverse content,and fast dissemination speed.They have expanded the scope of information dissemination,accelerated the speed of information dissemination,and met the needs of people in the era of fragmented information consumption to quickly obtain information.The massive user traffic has attracted competition from the advertising and e-commerce industries.Accurately predicting the future popularity of short videos and building a value evaluation system for short videos is the key to producing and selecting high-quality content.Currently,single-modal information processing capabilities cannot meet the complex and diverse prediction needs,and the industry lacks a multi-level value evaluation system based on market feedback.Therefore,this paper conducts an in-depth and meticulous study on the topic of short video popularity prediction and value evaluation based on deep learning technology,and proposes a popularity prediction model,builds a value evaluation system,and explores the factors affecting the evaluation results.Firstly,this study proposes a short video popularity prediction model based on deep learning multi-modal feature extraction technology and variational autoencoder.According to the time series variation law of short video playback volume,the study proposes a "short video popularity"measure standard.Considering the importance and correlation between the characteristic information and shared information of short videos,multimodal and multi-feature extraction are carried out for the visual,auditory,textual,and social modalities of short videos.The pain point caused by the lack of model reuse due to modal loss in previous studies is solved by using expert proíduct.After obtaining the hidden layer variable,a reliable time series sequence processing based on the long-short-term memory network is used in the decoding part.DenseNet structures and CNNs are added between the encoder and decoder to construct an autoencoder model,MASSL,which can alleviate the gradient disappearance problem and enhance the network translational invariance.The study also provides a dataset that is consistent with the short video popularity prediction scenario,and achieves advantages in comparison with the cutting-edge UGC content popularity prediction methods.Secondly,based on the short video popularity prediction results and market feedback,this study constructs a short video value evaluation system.Based on the dataset required for short video popularity prediction,the user behavior indicators are enriched and optimized after being stabilized,and the rationality of measuring short video market feedback based on user behavior indicators is demonstrated through KMO and Bartlett tests and principal component analysis.The expression of market feedback value is quantified reasonably through the factor loading matrix.Subsequently,from the time dimension,a short video value evaluation system is proposed.The study uses the multiple linear regression analysis method to conduct correlation analysis and regression coefficient calculation on the video value evaluation variable indicators,and further explores the positive or negative effects of different factors on market feedback.Based on this,corresponding improvement and optimization measures are proposed to improve the quality and value of short videos.Overall,this study proposed a popularity prediction model and a value evaluation system that combines multi-modal intelligent fusion,association mining,and structured prediction.This model has made an important contribution to the better application of deep learning in online video media services.