Combining different metadata views for better recommendation accuracy

Abstract Recommender systems emerged as means to help users deal with information overload by filtering content based on their preferences. Regardless of the recommendation method, there has been a recent interest in using user reviews as source of information, since they contain both detailed items’ descriptions as well as users’ opinions. Even though several works have been done in the subject, very few of them consider different views that the items may have towards their features, selecting only a method of weighting their features. In this work, we propose a system that combines two item representations that represent different views of the same feature set: one based on its statistics and the other based on its quality. Features are disambiguated concepts extracted from users’ reviews. We propose several strategies divided into three categories: pre combination, neighborhood combination and post combination. We evaluate our strategies in two data sets, comparing them with each other and against the isolated item representations, as well as a representation baseline based on terms and sentiment analysis. Results are promising showing that some combinations are capable of producing better rankings than their isolated versions.