Information-Theoretic Term Selection for New Item Recommendation

Recommender systems aim at predicting the preference of a user towards a given item (e.g., a movie, a song). For systems that must cope with continuously evolving item catalogs, there will be a considerable rate of new items for which no past preference is known that could otherwise inform preference-based recommendations. In contrast, pure content-based recommendations may suffer from noisy item descriptions. To overcome these problems, we propose an information-theoretic approach that exploits a taxonomy of categories associated with the cataloged items in order to select informative terms for an improved recommendation. Our experiments using two publicly available datasets attest the effectiveness of the proposed approach, which significantly outperforms state-of-the-art content-based recommenders from the literature.