Abstract: Nowadays, an increasing number of applications and services encourages users to express their emotions via images openly. Different from traditional visual sentiment classification, visual sentiment distribution learning is to explore the overall distribution to present the relative importance of sentiment labels. Considering most relevant studies that failed to completely model correlation structures or explicitly applied to unknown instances, in this paper, we proposed a low-rank latent Gaussian graphical model estimation (LGGME) method to solve visual sentiment distribution learning tasks. The main characteristics of LGGME are three folds: 1) an integrated inverse covariance matrix whose parameters characterize the latent correlation structures between and within features and sentiments is estimated with a sparse Gaussian graphical model; 2) a multivariate normal assumption is assigned on the concatenated latent feature representations and the estimated sentiment distributions instead of the original observations for a reasonable surrogate; 3) the latent feature representations are projected from a low-rank subspace which is also available for unseen instances and the estimated sentiment distributions are evaluated by KL divergence to ensure a suitable setting for distribution learning. We further developed an effective optimization algorithm based on the alternating direction method of multipliers (ADMM) for our objective function. Experiment results conduced on three publicly available datasets demonstrate the superiority of our proposed method.
0 Replies
Loading