Rating bias and preference acquisition

Personalized systems and recommender systems exploit implicitly and explicitly provided user information to address the needs and requirements of those using their services. User preference information, often in the form of interaction logs and ratings data, is used to identify similar users, whose opinions are leveraged to inform recommendations or to filter information. In this work we explore a different dimension of information trends in user bias and reasoning learned from ratings provided by users to a recommender system. Our work examines the characteristics of a dataset of 100,000 user ratings on a corpus of recipes, which illustrates stable user bias towards certain features of the recipes (cuisine type, key ingredient, and complexity). We exploit this knowledge to design and evaluate a personalized rating acquisition tool based on active learning, which leverages user biases in order to obtain ratings bearing high-value information and to reduce prediction errors with new users.