Reader Response Prediction and Cover Design Recommendation Based on Book Cover Elements

This study proposes a data-driven framework for book cover design that integrates prediction, generation, and user validation to examine the feasibility of evidence-based design recommendations. A dataset of 11, 195 Amazon books was constructed, from which ten design elements (color, layout, typography, and object type) were extracted. Machine learning models were tested to predict ratings and daily reviews. XGBoost performed best for rating prediction, while fine-tuned BERT yielded the lowest error for reviews. To examine whether model-selected cover candidates would align with actual reader responses, we conducted an online user study. Participants evaluated the highest- and lowest-predicted candidates on purchase behavioral intention. ANOVA results showed that review-based candidates significantly influenced purchase intention, with higher predicted reviews eliciting stronger purchase intentions. This study makes three contributions: (1) a reader-response prediction model, (2) an integrated prediction-generation framework, and (3) empirical validation via a user study. Overall, review signals proved more reliable than ratings in explaining purchasing behavior, offering practical implications for publishers and broader applications in visual media.