AUSTIN — Many music streaming services like Pandora and Spotify use a single song, artist, or genre choice to automatically generate playlists for users. One lone recommendation can end up creating a playlist that, for many, just grows plain boring after a few songs. Researchers from the University of Texas at Austin sought out to make the streaming experience more unique and enjoyable for users, so they created a “personalized DJ” that tailors playlists to each user’s mood.
The program they created, called “DJ-MC,” uses machine learning algorithms to create adaptive playlists for each individual user. They want their program to outperform the major music streaming services in the way recommendations are made to users.
“Whether you’re getting into the car after a long day of meetings, or you’re getting out of bed on a weekend morning, it should tailor its recommendations to your changing moods,” says Maytal Saar-Tsechansky, a professor of Information, Risk, and Operations Management at the McCombs School of Business, and senior author on the study, in a release.
The program is the product of the efforts of Saar-Tsechansky, UT Computer Science Professor Peter Stone, and computer science PhD candidate Elad Liebman. Liebman also has a degree in music composition, and he combined his interests in music and computer science to come up with the idea for the program.
DJ MC relies on feedback so it’s machine learning algorithm can create the optimal listening experience for users. The program plays a song and then asks the listener to rate the song. DJ MC then takes the rating under consideration when picking out the next songs to play. The program not only plans what songs to play next, but also the best order to play them in. This feature is what really distinguishes DJ MC from major music streaming services, since it makes it feel like playlists are curated by a deejay instead of randomly assembled.
As users listen to a song, DJ MC is busy generating tens of thousands of possible sequences for the next batch of songs to play. The program predicts which sequence the user will like the most, and selects the next song to play from that batch of songs. As the next song plays DJ MC repeats the process of generating songs and predicting the most desirable one.
This process of generating possible sequences and choosing the best one is called a “Monte Carlo search” learning algorithm. The research team took the initials of Monte Carlo to name their virtual DJ program.
The researchers suggest that this style of machine learning doesn’t need to be limited to music. A similar type of program can be made for anything that people consume. It all depends on the type of data that’s fed to the machine learning program.
“Learning algorithms don’t have taste, they just have data,” Liebman says. “You can replace the dataset with anything, as long as people are consuming it in a similar fashion.”
Saar-Tsechansky adds: “It can work in any case where you’re recommending things to humans, experienced in a sequence. It could even be food.”
The study is published in MIS Quarterly.