Machine Learning Methods for Music Discovery and Recommendation

In this talk I will relate current work at Google in music recommendation to the challenge of automatic music annotation ("autotagging"). I will spend most of the talk looking at (a) signal processing and sparse coding strategies for pulling relevant structure from audio, and (b) training multi-class ranking models in order to build good music similarity spaces. Although I will describe some technical aspects of autotagging and ranking via embedding, the main goal of the talk is to foster a better understanding of the real-world challenges we face in helping users find music they'll love. To this end I will play a number of audio demos illustrating what we can (and cannot) hope to achieve by working with audio.