Using an ensemble of features for personalized recommendations of scientific publications

Maintaining reviews of scientific publications as soon as new relevant publications are available is a typical challenge to many research communities. We address this challenge as a content-based recommendation problem, where the publications already selected for a review drive the recommendation of the new publications. In addition, resources such as domain databases, ontologies and academic graphs provide structured information about publications (e.g., authors, journals, conferences). Our experiments show that a simple model based on that structured information to represent publications achieve high precision and recall, and outperform models that use more sophisticated representations based on embeddings.