A sentiment-based item description approach for kNN collaborative filtering

Abstract

In this paper, we propose an approach based on sentiment analysis to describe items in a neighborhood-based collaborative filtering model. We use unstructured users' reviews to produce a vector-based representation that considers the overall sentiment of those reviews towards specific features. We propose and compare two different techniques to obtain and score such features from textual content, namely term-based and aspect-based feature extraction. Finally, our proposal is compared against structured metadata under the same recommendation algorithm, whose results show a significant improvement over the baselines.