Domain Adaptation for Emotional Score Prediction of Advertising Stock Assets

Abstract

In proportion of the digital ads market growth, the demand for systems that help graphic designers, such as asset image search engines, is increasing. Semantic labeling, e.g., emotions or impressions, to images using machine learning is one of the smart ways to provide such the system. However, due to the wide variety of artificial media expressions, the model learned by a specific medium will not work well to the other medium. In this paper, we propose an unsupervised emotional score prediction method for artificial images expressed by various media. Through experiments, we show the superiority of our method over the naive supervised approach.