Correcting noisy ratings in collaborative recommender systems

Recommender systems help users to find information that best fits their preferences and needs in an overloaded search space. Most recommender systems research has been focused on the accuracy improvement of recommendation algorithms. Despite this, recently new trends in recommender systems have become important research topics such as, cold start, group recommendations, context-aware recommendations, and natural noise. The concept of natural noise is related to the study and management of inconsistencies in datasets of users’ preferences used in recommender systems. In this paper a novel approach is proposed to detect and correct those inconsistent ratings that might bias recommendations, whose main advantage regarding previous proposals is that it uses only the current ratings in the dataset without needing any additional information. To do so, this proposal detects noisy ratings by characterizing items and users by their profiles, and then a strategy to fix these noisy ratings is carried out to increase the accuracy of such recommender systems. Finally a case study is developed to show the advantage of this proposal to deal with natural noise regarding previous methodologies.