Micro-Video Popularity Prediction Via Multimodal Variational Information Bottleneck

As an emerging type of user-generated content, micro-video drastically enriches people's entertainment experiences and social interactions. However, the popularity pattern individual still remains elusive among researchers. One major challenges is that potential a tends to fluctuate under impact various external factors, which makes it full uncertainties. In addition, since micro-videos are mainly uploaded by individuals lack professional techniques, multiple types noise could exist obscure useful information. this paper, we propose multimodal variational encoder-decoder (MMVED) framework for prediction tasks. MMVED learns stochastic Gaussian embedding informative its level while preserves inherent uncertainties simultaneously. Moreover, through optimization deep information bottleneck lower-bound (IBLBO), learned hidden representation shown be maximally expressive about target compressive in features. Furthermore, Bayesian product-of-experts principle applied encoder, where decision keeping or discarding made comprehensively with all available modalities. Extensive experiments conducted on public dataset collect from Xigua demonstrate effectiveness proposed framework.