Taxonomy-driven content-based recommendation for new itens

Abstract

Recommender systems aim at predicting the preference of a user towards a given item such as a movie, a song, or a news story. Effective recommendations can be produced through collaborative filtering, in which case the previously manifested preferences of a community of users are leveraged to inform the recommender system. Effective recommender systems must cope with an evolving item catalog and an increasing user base, leading to a considerable rate of new items and new users, both with unknown past preferences. This problem, known as the cold-start recommendation problem, may hamper the performance of recommender systems that are based solely on collaborative filtering. To overcome this problem, we propose an approach that exploits content-based features derived from taxonomies associated with the cataloged items. In contrast to previous content-based recommendation approaches, our approach is domain-agnostic, and can be directly deployed to produce effective cold-start recommendations in different domains. For domains where an explicit taxonomy is not available, we show that a suitable one can be derived implicitly using Latent Dirichlet Allocation. Our experiments using two publicly available datasets with distinct levels of sparsity attest the effectiveness of the proposed approach, which significantly outperforms several state-of-the-art baselines from the literature.