Music catalogs in music streaming services, on-line music shops and private collections become increasingly larger and consequently difficult to navigate. Music recommender systems are technologies devised to support users accessing such large catalogs by automatically identifying and suggesting music that may interest them. This thesis focuses on the machine learning aspects of music recommendation with contributions at the intersection of recommender systems and music information retrieval: I investigate and propose recommender systems that observe and exploit the particularities of the music domain. The thesis specializes in "hybrid" music recommender systems, so called because they combine two fundamentally different types of data: (1) user-music interaction histories (e.g., the music that users recently listened to, or "liked"), with (2) descriptions of the musical content (e.g., the genre, or acoustical properties of a song). The proposed hybrid music recommender systems integrate the strengths of these two types of data into enhanced standalone systems. This is in contrast to most previous approaches in the literature, where hybridization was achieved through the heuristic combination of music recommendations issued by independent systems. The proposed hybrid music recommender systems are thoroughly evaluated against competitive recommender system baselines, for different music recommendation tasks, and on different datasets. According to the conducted experiments, the proposed systems predict music recommendations comparably or more accurately than the considered baselines, with the improvements being largely explained by their superior ability to handle infrequent music items. In this way, the proposed hybrid music recommender systems provide means to alleviate the so-called "cold-start" problem for new releases and infrequent music and enable the discovery of music beyond the charts of popular music. Special attention is paid to the particularities of the music domain. I focus on two important music recommendation tasks: music artist recommendation, focusing on general, stable user music preferences, and music playlist continuation, focusing on local relationships in short listening sessions. I exploit data sources abundant in the context of on-line music consumption: user listening histories, hand-curated music playlists, music audio signal, and social tags. I investigate challenges specific to modeling music playlists: the choice and the arrangement of songs within playlists, and the effectiveness of different types of music descriptions to identify songs that fit well together.