Can book covers help predict bestsellers using machine learning approaches?

As the book publishing market changes from offline to online, readers tend to purchase books while paying more attention to book covers and metadata rather than the actual book contents. We examine whether publishers can know users' satisfaction with books in advance, and both metadata and book covers help predict this satisfaction. Exploring effects of metadata and book covers on the satisfaction is not only necessary for publishers' perspectives, but also for librarians' perceptions. However, the majority of prior research on user preference-based book recommen-dation systems in both book industry and library system employed review comments, ratings, or book loan records. Thus, we open up the potentiality of other factors, which implicitly affect the satisfaction with books. We collected book titles, authors, publishers, reviews, ratings, and covers from the "Literature and Fiction" genre in the Amazon bookstore and conducted an experiment to predict readers' satisfaction ratings based on book reviews, metadata, and book covers. Several deep learning classifiers (CNN, ResNet, LSTM, BiLSTM, GRU, BiGRU) were employed. Reviews alone can reach a certain level of prediction performance, but adding metadata, cover images, and cover objects to a review-based predictive model slightly improves that performance. Based on these results, we confirmed that both metadata and book covers improve predicting readers' perceived satisfaction. This study is a pilot exploration of the idea that multimodal approaches can improve the prediction of the perceived satisfaction of book readers. Moreover, we have publicly released both source codes and data samples employed in this study.