Recommender systems are ubiquitous on the Web, improving user satisfaction and experience by providing personalized suggestions of items they might like. In the past years, knowledge-aware recommender systems have shown to generate high-quality recommendations, combining the best of content-based and collaborative filtering. The crucial point to leverage knowledge graphs to generate item recommendations is to be able to define effective features for the recommendation problem.
This thesis investigates the use of knowledge graph embeddings for recommender systems. Knowledge graph embeddings learn a mapping from the knowledge graph to a feature space solving an optimization problem, minimizing the time-consuming endeavor of feature engineering and leading to higher quality features. We introduce entity2rec, which learns user-item relatedness for item recommendation through property-specific knowledge graph embeddings. entity2rec has been benchmarked with a set of existing knowledge graph embeddings algorithms (translational models, node2vec) that we have applied to the recommendation problem and with popular collaborative filtering and hybrid algorithms on three standard datasets. entity2rec has shown to generate accurate and non-obvious recommendations, achieving high accuracy, serendipity, and novelty, and to be particularly effective when the dataset is sparse and has a low popularity bias. Furthermore, entity2rec is based on a recommendation model that encodes the semantics of the knowledge graph and can thus be interpreted and configured for a particular recommendation problem. entity2rec has also been tested in a cold start scenario with real new users through a web application called Tinderbook. Tinderbook is a web application that recommends books to users, given a single book that they like, leveraging an item-item relatedness measure based on entity2rec.