Abstract
Recommender systems allow users to express preferences for items using a rating scale. However, employing this scale differs from one user to another, and sometimes it seems that each user has a personal scale. Hence it is vital to capture the user preferences for the rating scale for better recommendations. Previous work tried to weight either the similarity value or the individual ratings of items. However, weighting the rating scale itself to reflect personal user interests is more effective. This reduces the problem space by directly targeting the tool used to express user preferences, not individual ratings for items. This paper proposes different approaches for generating a personalized rating scale that captures the user preferences reflected by the statistical properties of his profile. Some approaches consider the features of the active user, while others also consider the attributes of the training user. The experimental results disclose different user behavior patterns for the rating scale. They show that the system performance is improved so much by introducing weights to the rating scale with 70% improvement for some schemes.