Recommender systems help users to find information that best fits their preferences and needs in an overloaded search space. Most of recommender systems research focuses on improving recommendation methods to obtain a higher accuracy in recommendations. However, the study of user's inconsistencies, so-called natural noise, is becoming a hot topic in Recommender Systems. In this contribution is proposed a novel approach to detect and correct those inconsistent ratings that might bias recommendations, by using global information about user and item preferences. This proposal characterizes items and users by their ratings and classifies a rating as noisy if it contradicts user or item tendencies. This approach just utilizes ratings on the contrary of previous proposals that use additional information like item attributes or user interaction.