Predicting the popularity of micro-videos via a feature-discrimination transductive model

Abstract

Nowadays, with the development of social media networks, micro-videos, an emerging form of user-generated contents (UGCs), are gradually attracting greater interest. Some of them are widely spread, while others draw little attention. The popular micro-videos have significant commercial potential in many ways, such as online advertising and bandwidth allocation. In recent years, the popularity prediction of long videos, web images and texts have gained abundant theoretical support and made great practical success. However, little research has been conducted on micro-videos. There are three difficulties in dealing with the problem: (1) micro-videos are short in duration; (2) the quality of micro-videos is relatively poor; (3) micro-videos can be described by multiple heterogeneous features involving social, visual, acoustic and textual modalities. For these purposes, we presented a feature-discrimination transductive model (FDTM). The proposed method regards the multi-view features as two properties: the low-level features and the attribute features. We divided the micro-videos into different levels of popularity via the attribute features and predicted the popularity scores via the low-level features precisely. Moreover, in the process of prediction, we sought a latent common feature subspace, where the micro-videos can be comprehensively represented. The latent subspace can aggregate the multiple low-level feature information to alleviate the problem of information insufficiency. Extensive experiments on a public dataset show that the proposed method achieves significant improvements compared with the best-known models.