Prediction of the emotions and preference for movie poster designs based on graphical features: A machine learning-driven approach

Abstract

Advances in data science have allowed us to investigate human emotion and preference responses based on the learning of the image data set. However, to date, studies have been limited to photography, meaning that the practice of graphic design remains unexplored. We conveyed a workshop to exploit graphic elements that are perceptually relevant. By identifying and quantifying the visual elements of poster designs, we defined the machine learning features. We collected emotion and preference responses from participants with regard to pleasure, arousal, dominance, and preference using a Likert scale, and combined them with the features to build a dataset. The responses were carried out twice to form a baseline of accuracy. Applying machine learning methods such as CART, Random Forest, and XGBoost, we developed the models that predict individuals' emotions and preference responses. Human responses were repeated in the range of 62.81–87.50%, and the accuracy of the prediction model achieved between 53.38–83.20% using XGBoost. We discussed the prediction results with the participants and improved the features by adopting their opinions. With the new features, we conducted an additional survey to 15 people. The prediction performance increased by a maximum of 7.22%, depending on personalized models. This study demonstrates that a machine learning-driven modeling may predict one's emotions and preference based on graphical elements. Overall, the prediction performance was relatively satisfied compared to human repeated responses. Limitations and further investigations are discussed to obtain a more accurate and insightful estimation.