Deep low-rank matrix factorization with latent correlation estimation for micro-video multi-label classification

Currently, micro-videos are becoming an increasingly prevailing form of user-generated contents (UGCs) on various social platforms. Several studies have been conducted to explore the semantics of micro-videos and the behavior of individuals for various tasks, such as venue categorization, popularity prediction, and personalized recommendation. However, few studies have been dedicated to solving micro-video multi-label classification. More importantly, learning intrinsic and robust feature representations for micro videos is still a complicated and challenging problem. In this paper, we propose a deep matrix factorization with latent correlation estimation (DMFLCE) for micro-video multi label classification. In DMFLCE, we develop a deep matrix factorization component constrained by a low-rank constraint to learn the lowest-rank representations for micro videos and the intrinsic characterizations for latent attributes simultaneously. To explicitly exhibit the dependencies of the learned latent attributes and labels for improved classification performance, we construct two inverse covariance estimation components to automatically encode correlation patterns with respect to the latent attributes and labels. Experiments conducted on a publicly available large-scale micro-video dataset demonstrate the effectiveness of our proposed method compared with state-of-the-art methods. (c) 2021 Elsevier Inc. All rights reserved.