Movie Rating Estimation Based on Weakly Supervised Multi-modal Latent Variable Model

Abstract

In this paper, we present a method to estimate movie ratings based on a weakly supervised multi-modal latent variable model. One movie has multiple movie features (e.g., key-frames and descriptions) and one rating annotated by a user. Features included in one movie are not one-to-one correspondence with the rating, and general latent variable models cannot calculate latent variables under such a multiple-instance condition. To solve this problem, we propose a weakly supervised multi-modal label dequantized GPLVM (WmLDGP). WmLDGP can calculate latent variables by estimating label features for each scene based on a label dequantization scheme. The main contribution is introduction of the label dequantization scheme to the multiple-instance condition. Experimental results show the effectiveness of our model.