Reading Time: 4 minutes
Written by: Amy Shaw
The methods for recommending music to listeners have gradually shifted over time due to advances in technology. Through the years, most recommendations have been focussed around a user’s listening history alone. This has proved ineffective as a method overall and is an extremely limited approach in ascertaining a user’s preferences. Currently, many more methods for recommending music are being brought to the surface through thorough research. These methods could shape the way listeners consume music through streaming platforms in the future. It is imperative for artists to be aware of the recommendation methods streaming services use. We’ve taken a look at the current and future state of music recommendation and how this is set to affect artists around the world.
Listening history
While someone’s listening history can provide a degree of insight into their taste, it is by no means ideal. This method involves looking at a user’s listening history and comparing them with other users. The basis is, if two listeners enjoy the same song, they will likely share the same taste overall, and will both enjoy the same tracks. Using a listener’s history does not in any way use any information about the tracks themselves, and relies purely on listening patterns, making it extremely limited. Recommending music in this way can make recommendations seem repetitive and can put much of the focus on the most popular tracks rather than up and coming artists. Brand new songs that are yet to be discovered are unlikely to be recommended at all when using this method, and many users prefer to discover new tracks rather than those that are currently performing well in the mainstream chart. The recommendations are highly predictable and will leave users bored and uninspired by their music library.
Content recommendation
Recommendations can also be based on the content of a track. Spotify has been taking strides in basing its recommendations on data, including tags, artist info, lyrics, and recent reviews and interviews. As well as this, Spotify has been focussing on basing recommendations on a track’s audio signal. Using a track’s audio signal it is easy to determine genre and instruments used, but not so easy to ascertain listener preference. This is a much more effective method to measure user preference as it looks at the song itself rather than user data. Even though this type of recommendation is still extremely limited, it is still a stride in the right direction. It marks the first step into looking at tracks themselves when recommending music, something which looking at listening history was unable to provide.
Deep learning
A paper published at NIPS explored the possibility of recommending music through deep learning. Using audio signals to predict latent representations, this method can predict the representation of a song. This approach projects listeners and songs into a latent space, and the position of songs in the space encodes data around listening preferences. Simply put, if tracks are close together on the diagram, they share many similarities. If a track is close to a user, the listener is likely to enjoy it. Through mapping music in this way, it is possible to provide more accurate recommendations than ever before, a ground-breaking shift for music streaming services like Spotify and Apple Music. Using the t-SNE algorithm it is possible to map out tracks and cluster similar styles together. Using this method, particular genres cluster together, and the music industry is effectively mapped out completely.
Ripple effect
Although these advances in recommendation technology may seem more relevant to listeners than artists, this is not strictly true. The older methods of recommendation made sure that only popular tracks were being recommended rather than new releases that hadn’t yet amassed a high volume of streams. With these new methods, tracks have even more potential as soon as they are released. By using new methods, your tracks will be recommended to exactly the right people as early as possible, giving your tracks an even bigger opportunity for instant success. Up and coming artists will have just as much chance of success as popular artists, as tracks will not need to have already built up thousands of streams to stand a chance of being recommended. Overall, higher quality tracks that are preferable to a larger volume of people will stand the best chance of success on streaming platforms.
The gradual improvement of recommendation technology is positive for both artists and listeners. Listeners will be recommended tracks that they are more likely to enjoy and artists will stand more chance of being recommended to those who are most likely to show support. Artists’ tracks will also start being recommended as soon as they are released, allowing streams to begin building much more quickly than previously. This development will be positive for the music industry as a whole, prioritising quality and taste over artists who simply have a high volume of listeners. It is impossible to chart the full impact of this shift, but what we do know is that it will gradually shift the consumption of music using streaming services. The technology available is becoming more intelligent at a ground-breaking rate and is set to have a massive impact on the recommendations streaming services provide. As technology improves, the music industry looks set to benefit, and in turn, artists and listeners will also experience the positives. It is difficult to predict the future, but with streaming services becoming more intelligent in their software, the future is definitely set to be shaped greatly by recommendations provided to listeners.
Original:
https://moonjelly.agency/recommending-music-to-listeners-using-deep-learning/
Comments