Prediction of individual preference for movie poster designs based on graphic elements using machine learning classification

Abstract

We attempted to predict an individual’s preferences for movie posters using a machine-learning algorithm based on the posters’ graphic elements. We transformed perceptually essential graphic elements into features for machine learning computation. Fifteen university students participated in a survey designed to assess their movie poster designs (Nposter = 619). Based on the movie posters’ feature information and participants’ judgments, we modeled individual algorithms using an XGBoost classifier. We achieved prediction accuracies for these individual models that ranged between 44.70 and 71.70%, while the repeated human judgments ranged between 61.90 and 87.50%. We discussed technical challenges to advance prediction algorithm and summarized reflections on using machine learning-driven algorithms in creative work.