The ultimate mission of a Recommender System (RS) is to help users discover items they might be interested in. In order to be really useful for the end-user, Content-based (CB) RSs need both to harvest as much information as possible about such items and to effectively handle it. The boom of Linked Open Data (LOD) datasets with their huge amount of semantically interrelated data is thus a great opportunity for boosting CB-RSs. In this paper we present a CB-RS that leverages LOD and profits from a neighborhood-based graph kernel. The proposed kernel is able to compute semantic item similarities by matching their local neighborhood graphs. Experimental evaluation on the Movielens dataset shows that the proposed approach outperforms in terms of accuracy and novelty other competitive approaches.