Music Recommendation System Using the Million Song Dataset

We attempted to build a recommendation system to recommend songs to users using data from the Million Song Dataset available on Kaggle. The main focus of this paper was to use collaborative filtering algorithms that only utilize user feedback in order to predict what songs users may like. For benchmarking the algorithms, we used the same Mean Average Precision score truncated at 500 recommended songs that was used in the original Kaggle competition. It was discovered that probabilistic matrix factorization with a MAP value of 0.014 did not improve results much from using a baseline of simply recommending popular songs, while artist-based popularity along with user-based and item-based collaborative filtering methods yielded much better results with the best method giving a MAP value of 0.048.