结合多模态时序特征与变分编码的视频流行度预测算法OA

Currently,video popularity prediction is widely used in social media marketing,intelligent advertisement place-ment,etc.Traditional prediction methods based on multiple linear regression,support vector regression model,etc.,have some limitations in dealing with data diversity and capturing prediction uncertainty.A multi-modal temporal variational autoencoder(MTVAE)video popularity prediction model is proposed from multi-modal data.Multi-modal features such as visual,audio,text and social are extracted from video data,and the MTVAE is used to downsize the multi-modal fea-tures and extract the hidden information.The intrinsic connection of long sequential features is mined with the help of temporal convolution network,while residual connectivity is added at the network level to enhance the deep expression capability and the training stability of the model.A Bayesian expert product system is applied to perform multi-modal fea-ture fusion and prediction discrimination.In addition,whole-process multi-attribute video popularity prediction dataset(WPMAD)based on Sina Weibo platform is constructed,which makes up for the deficiencies of the current dataset in terms of modal diversity and temporal integrity.A large number of experimental results show that the MTVAE model improves prediction accuracy by 3.93 percentage points on the WPMAD dataset and also improves prediction accuracy on the public MicroLens dataset.Compared with existing prediction methods,it performs better and provides innovative tech-nical support for applications such as content recommendation and public opinion monitoring on social media platforms.