At Spotify the main search mode is search-as-you-type where new results are presented after each character entered by the user. These partial search queries are inherently very ambiguous. For instance, when typing “e”, the user might be looking for “Elvis”, “Eminem” or something else. Without information about the user we have to resort to sorting artists based on popularity. However, we can do better. By mining user data and training a machine learning model to predict clickthrough logs, we have built a ranking model that personalizes the order of presented items and as a result achieves higher total success ratios in search sessions.
In this talk Magnus gives an overview of the Learning to Rank problem and present some details of the current implementation for search at Spotify.