Recommender systems help users to deal with the information overload problem by producing personalized content according to their interests. Beyond the traditional recommender strategies, there is a growing effort to incorporate users' reviews into the recommendation process, since they provide a rich set of information regarding both items' features and users' preferences. This article proposes a recommender system that uses users' reviews to produce items' representations that are based on the overall sentiment toward the items' features. We focus on exploiting the impact that different feature extraction techniques, allied with sentiment analysis, cause in an item attribute-aware neighborhood-based recommender algorithm. We compare four techniques of different granularities (terms and aspects) in two recommendation scenarios (rating prediction and item recommendation) and elect the most promising technique. We also compare our techniques with traditional structured metadata constructions, which are used as the baseline in our experimental evaluation. The results show that the techniques based on terms provide better results, since they produce a larger set of features, hence detailing better the items.

In this paper, we propose a technique to automatically describe items based on users' reviews in order to be used by recommender systems. For that, we extract items' features using a robust term extraction method that applies transductive semi-supervised learning to automatically identify aspects that represent the different subjects of the reviews. Then, we apply sentiment analysis in a sentence level to indicate the polarities, yielding a consensus of users regarding the features of items. Our approach is evaluated using a collaborative filtering method, and comparisons using structured metadata as baselines show promising results.

No abstract

No abstract