Why Do CF and Review Models Disagree? User-Specific Rating-Review Scale Gaps

초록

Ratings and reviews capture different aspects of preference, so rating-based collaborative-filtering and review-based models may disagree on a user-item interaction. Aggregate accuracy does not reveal why they disagree or which user characteristics are associated with it. We examine whether this reflects discrepancies between users' rating tendencies and review tone. We introduce the rating-review scale gap (RRS gap), the standardized difference between a user's rating leniency and review-tone bias. Large disagreement alone does not identify the better model, whereas the RRS gap is consistently associated with whether the review-based model predicts above or below the rating-based model, with domain-dependent strength. Such user-specific discrepancies may therefore help explain model disagreement. Our source code is available at https://github.com/alsrb9936/rating_and_reviews. © 2026 Copyright held by the owner/author(s).

키워드