A Matrix Factorization Algorithm for Music Recommendation using Implicit User Feedback

The goal of recommender systems is to make personalized product recommendations based on users’ taste. As the Netflix challenge demonstrated, one of the the most effective way to build such systems is through matrix factorization. Matrix factorization algorithms utilize prior product feedback given by users to automatically build user and product profiles. A product can then be recommended to a user if the user’s profile closely matches that of the product. Unfortunately, most of the research in matrix factorization focuses on explicit feedback datasets, where users make their preferences known by directly rating subsets of available products on a fixed scale. In many real-worlds applications, however, such direct ratings are unavailable. Instead, implicit feedback must be used, such as browsing history and view counts. In this work we present a recommendation algorithm which uses implicit play frequency information to recommend artists to the users of the Last.fm1 music website. We demonstrate how an existing algorithm for explicit feedback datasets can be adapted to work with implicit data. We further evaluate the performance of our recommendation algorithm on a Last.fm dataset, investigating the effects of regularization, normalization and the number of features on the quality of recommendations. We also discuss the viability of our approach in a real-world setting. http://www.last.fm