Multi-modal Variational Auto-Encoder Model for Micro-video Popularity Prediction
Zhuoran Zhang, Shibiao Xu, Li Guo, Wenke Lian
Abstract
Popularity prediction of micro videos on multimedia is a hotly studied topic due to the widespread use of video upload sharing services. It’s also a challenging task because popular pattern is affected by multiple factors and is hard to be modeled. The goal of this paper is to use feature extraction techniques and variation auto-encoder (VAE) framework to predict the popularity of online micro-videos. First, we identify four declarable modalities that are important for adaptability and expansibility. Then, we design a multi-modal based VAE regression model (MASSL) to exploit the domestic and foreign information extracted from heterogeneous features. The model can be applied to large-scale multimedia platforms, even the modality absence scenarios. With extensive experiments conducted on the dataset, which was originally generated from the most popular video-sharing website in China, the result demonstrates the effectiveness of our proposed model by comparing with baseline approaches.