Media recommendation using internet media stream modeling

摘要

Media item recommendations, such as music track recommendations, may be made using one or more models generated using data collected from a plurality of media stream sources, such as, for example, Internet radio stations. In an initial, bootstrapping phase, data about media items and media stream playlists of media stream sources may be used to generate a model, which comprises latent factor vectors, or learned profiles, of media items, e.g., tracks, artists, etc. Such a bootstrapping phase may be performed without user data, such as user playlists and/or user feedback, to generate a model that may be used to make media item recommendations. As user data becomes available, e.g., as users of a recommendation service provide user data, the user data may be used to supplement and/or update the model and/or to create user profiles.

主权项

1. A method comprising:
collecting, via at least one computing device, a training data set comprising data about a plurality of playlists of a plurality of media streams from a number of Internet streaming media stations, for each occurrence of a media item in a playlist, the training data set comprising information identifying the media item, the playlist, at least one artist associated with the media item, and a timestamp indicating a play time of the media item; generating, via the at least one computing device, at least one latent factor model using the training data set, the at least one latent factor model modeling a plurality of media items and the plurality of media streams, the at least one latent factor model is at least initially generated without using user data identifying user media item preferences, the training data at least initially being used in place of the user data to generate the at least one latent factor model; and using, via the at least one computing device, the at least one latent factor model to make a number of recommendations, at least one recommendation of the number being made using the at least one latent factor model generated without using the user data.