With the prevalence of Internet and mobile technology, it becomes much easier for people to access millions of songs anywhere, anytime. The explosive increase of music collections on the Internet makes music retrieval much harder and more complicated for normal users. A music recommender system is a music retrieval system, which actively recommends music collections to users based on their preferences.
Many general recommender system techniques are also suitable for music recommendation. In this work, we evaluated five music recommendation algorithms: popularity-based, item-based, userbased, Pure-SVD, and incremental-SVD approach. The experiment was conducted on a real-world music listening history dataset from Last.fm. We used recall@N to evaluate the performance of different recommendation algorithms on a top-N recommendation task.
The result showed that the SVD-based approaches generally have better performance than other types of recommendation algorithm. The incremental-SVD approach produced the best performance among all, and the Pure-SVD approach achieved the second-best result with very good scalability.
