Cold-start music recommendation using a hybrid representation.

Digital music systems are a new and exciting way to dis- cover, share, and listen to new music. Their success is so great, that digital downloads are now included alongside tra- ditional record sales in many o cial music charts [10]. In the past listeners would rely on magazine, radio, and friends reviews to decide on the music they listen to and purchase. In the internet age, this style of nding music is being su- perseded by music recommender systems. The shift from listening to hard copies of music, such as CDs, to online copies like MP3s, presents the interesting new challenge of how to recommend music to a listener. In such recommender systems, a user will typically provide a track that they like as a query, often implicitly as they listen to the track. The system must then provide a list of further tracks that the user will want to listen to. Many websites exist that provide such recommender systems, and many of the systems provide very good recommendations. However, there are still scenarios that these systems struggle to han- dle, and where recommendations can be unreliable. Online music systems allow users to tag any track with a free-text description. A recommender system can then determine the similarity between tracks based on these tags, and make recommendations. However, when a track is new to the system it will have no tags. This means that the track is never recommended, and in turn, the track is very unlikely to be tagged. Turnbull et. al [11] show that social tags tend to be very sparse, and that a huge popularity bias exists. This is further con rmed by data released by Last.fm [7] as part of the million song dataset [3]: from a vocabulary of over 500000 tags, each track, on average, has only 17 tags; 46% of tracks have no tags at all. This scenario is often referred to as the cold-start prob- lem; the results of which means large volumes of music are excluded from recommendations, even if they may be an excellent recommendation. The aim of our hybrid repre- sentation is to reduce the e ects of the cold-start problem, therefore increasing the recommendation quality of the over- all system