Behind the precise recommendations is a powerful algorithm.
The Translation Bureau is a compilation team that focuses on technology, business, workplace, life and other fields, focusing on foreign new technologies, new ideas, and new trends.
Editor’s note: The main difference between Spotify and other music software is that there is a unique algorithm that accurately recommends songs that match your taste, tailoring your own listening experience. This article is translated from Medium, author Dave Gershgorn, originally titled “How Spotify’s Algorithm Knows Exactly What You Want to Listen To”, I hope to inspire you.
Photo: Aytac Unal/Anadolu Agency/Getty
Spotify is doing everything it can to make users hear more music. The company has developed an algorithm to manage software, from personal homepages to customized song lists, such as Discover Weekly, and continues to experiment with new ways to understand music and understand why people are I like listening to a song or a genre of music.
When its competitors, such as Apple Music, Amazon Prime Music, and Google Music, also rely on paid playlists created by paying users and communities, Sound Field The biggest difference with them is to provide customers with a large number of personalized music and provide extensive knowledge of music. Sound Field also needs to continue to develop better algorithms because it is the only way to tailor the music experience to each of more than 200 million users. As Sonic strives to expand its business, the above-mentioned factors that make it unique need to be attractive enough for consumers to subscribe to the service.
A typical example of how the algorithm manages the listening experience is the home screen of the sound field application. Mounia Lalmas-Roelleke, head of research and development, said in a speech at a webinar earlier this year that Sound Field’s goal is to help users quickly find the music they like.
Monia explained, the main pageControlled by an artificial intelligence system called BaRT (Bandits for Recommendations as Treatments, “There are only robbers who recommend music for one thing”). The task of this system is to organize each user’s homepage in a personalized way, including “music racks”, including a series of songs of the same theme, such as “artist best” or “ambient music”, then the artificial intelligence The system will make the song list appear on the music shelf of the corresponding theme.
The BaRT system is the core means of all aspects of sound field balance. It has only one purpose, based on the music you listened to earlier, giving users the music that Spotify believes users will like. But Sumada must also add new music to it so that you don’t fall into the loop of listening to the same kind of music.
The use of BaRT can be attributed to two concepts: deep digging and exploration. When Sound Field chooses the “deep digging” mode, it uses the collected user information, considering your music listening history, the songs you skipped, the playlists you created, what you did with the social features of the platform, Even including your location. But when Sound Field chose the “Explore” mode, it used outside information, such as playlists and artists that matched your musical taste but you haven’t heard before, as well as the heat of other artists.
As important as Sound Field’s ability to dig deep and explore is how the app explains its push options to the user. Every tab on the music rack, like “jump back here” or “more content you like”, tells the user why these playlists are recommended. According to a 2018 research paper on BaRT, Sound Field found that interpretation is critical to gaining user trust.
The success of BaRT is measured by whether you listen to the music on the shelf and how long it has been listening. If a song is played for more than 30 seconds, the algorithm will track the process to remember this recommendation as correct. The longer you listen to the recommended playlist or collection, the better the system’s recommendations are considered to be.
Spotify seems to be determining whether a person likes a song for a comfortable time of 30 seconds. In an interview with the news media Quartz in 2015, Soundfield’s product director Matthew Ogle mentioned at 30Skip songs before seconds is equivalent to negating the “weekly discovery” playlist.
The company clearly stated in its research that in order for these algorithmic services to be successful, each operation of the user while using the service must be tracked and recorded.
There is not much academic work to provide a detailed and complete introduction to the mechanism for “weekly discovery” playlists. In the Quartz report, Ogle outlined the system, and this overview is related to a 2015 report made by Spotify employees, which is slightly more technical. “Weekly Discovery” is a playlist of 30 songs from other users whose music preferences are similar to yours, recent music blog posts, and other songs that sound like your favorite music. In 2014, Sound Field spent $100 million to acquire the startup “The Echo Nest” to improve the quality of recommendations. Eric’s co-founder Brian Whitman wrote in 2012 that his software needs to search more than 10 million music-related web pages every day to understand the trends in the music world.
Whitman said: “Our system searches every word on the Internet about every word of music, looking for descriptive words, noun phrases, and other texts.”
In 2014, Sander Dieleman worked as an intern at Sound Field and did some basic work on analyzing musical auditory similarities. He explained audio analysis algorithms in his personal blog. The initial problem was that new music was uploaded to the sound field every day, but if the music was not the work of a previously popular artist, there would be no system to recommend it. When no one knew the artist at the beginning, the collaborative filtering algorithm (a method of recommending music that they liked to people with similar musical interests) did not work.
Dillerman called this a “Cold-start problem.”
The solution is to analyze the audio itself and train an algorithm to learn to identify different possible sources of musical attraction. Some of Diliman’s experiments identified specific sources of attraction for songs, such as distorted guitars, while others identified more abstract concepts, such as genres.
This method has now become an important part of the “Weekly Discovery” playlist, which is why you see artists who have never heard of it appear on the list of recommendations.
The
algorithm can be applied to all aspects of sound field software. There are not only recommended algorithms for “weekly discovery” and features such as the home screen, but also some gadgets that you may have used but never realize that they are the product of relatively sophisticated artificial intelligence research.
Take an automatic continuous playlist as an example. This oneThe function analyzes the songs in a particular playlist and tries to predict which one is being played next, as if the person who created the list is constantly adding music to it. Sound Field hopes to think about how to build this feature in a new way, so it released a user-generated “Million Playlist Dataset” to understand what features a good set of tracks in a person’s mind should have. The company also invited other artificial intelligence researchers to try and solve the problem and publish the solution at the 2018 industry conference. Analysis of the game organizers after the game showed that more than 100 academic and industry teams were formed around the project. (We don’t know if the winner’s idea is really included in the sound field.)
Soundfield researchers have also been researching ways to detect different versions of songs. Because other versions may replace the original version you really want to hear out. The results of their final work can be highly accurate in distinguishing between original tracks and other versions, especially instrumental performances and live performances. Jazz is even more tricky because there are a lot of improvisations.
The team is also committed to aligning lyrics and singing moments, which not only helps the company develop the ability to display lyrics next to popular songs, but also provides new opportunities for Sound Field.
“Ordering lyrics by time makes possible applications such as karaoke, text-based song retrieval, and internal navigation of songs, enriching the user’s music experience,” Sonoda’s computer scientist wrote earlier this year.
In addition to the platform-oriented, Sound Field’s research also includes user-oriented. According to a study published in April 2019, Sound Field analyzed data from more than 16 million users and tracked their listening patterns from December 2016 to February 2018, including the number of times a particular artist or a particular song was played each day. And user’s location information, etc.
The above data, together with the gender and age of the user, allows Sonata to study whether a person’s musical taste changes after moving to another state, and how age affects the type of music a person likes.
The research team intuitively inferred the user’s residence in a subtle way: According to the location data, some users of the sound field will have three main holidays during the selected data period – 2016 Christmas, 2017 Thanksgiving And Christmas 2017The two holidays in the festival went to other states, and the research team speculated that the state they traveled to was their hometown.
After studying the musical tastes of the users of each state and then comparing those who have moved to different regions of the music trend, Sonoda’s team came to the conclusion that staying in a place for a long time does affect people’s Musical taste.
“Relocation does make a person’s musical taste slightly shift to the new environment. But the impact is minimal and the audience is more inclined to imitate their past,” they wrote.
The team also found that the music that users like during the 10-20 years of age will be his mainstream music in the future, which forms their “music personality.”
All this implies that in order to maintain its competitive advantage, Sound Field needs to continue to use mobile information from users. In the 2015 “Weekly Discovery” demo, Sound Field recorded 1 trillion bytes of user data every day.
However, the data is clearly for the user. The company clearly emphasized in the research that all of these algorithmic services that want to succeed must track and record every action the user performs on the software.
This is probably the secret weapon of music streaming. After all, the music industry giants, including Apple, which has tens of millions of paying users, have been struggling with Sumida, but it is still thriving.
Translator: Jane
Recommended reading: Why does science advise you to take a nap?