If you don’t like the tracks automatically offered to you by the web, it might not be you – it might be the AI.
At least some recommendation algorithms just aren’t very good at suggesting music to fans of hard rock and hip hop, according to a study by machine-learning experts and data scientists. The problem appears to be this, we’re told:
- People who prefer so-called hard music, which covers a wide set of genres from hard rock and punk to hip hop, aren’t much interested in music outside of their niche.
- People who prefer softer music, which is a limited set of genres, are happy to listen to artists outside of their niche.
- Algorithms are better at recommending tracks for easygoing ambient fans than for choosy hard music fans, or in other words, worse at picking tracks for hard music fans than for ambient lovers.
In short, hard music fans are picky and thus difficult to please, and algorithms can’t cope with that. That might seem an obvious conclusion but bear in mind you’d expect recommendation engines to overcome that and identify tracks people would enjoy regardless of taste. In reality, no dice, sadly.
“Some subgroups, like ambient listeners, seem to be more open to listening to music from other subgroups; plus, they are more similar to each other – all of this is great for recommendation algorithms and such users more likely accept recommendations from different groups,” Elizabeth Lex – co-author of this research, published in EPJ Data Science this month – told The Register.
“In contrast, hard rock and hip-hop low-mainstream listeners are, in our data, the least open to music of other subgroups, and within themselves, much more diverse, and thus, harder to satisfy with recommendations.”
The researchers, led by boffins at the Graz University of Technology (TUG) in Austria, analysed the performance of multiple recommendation engines on the music listening histories of about 4,000 people scraped from Last.fm users; the code involved is on GitHub, here.
Sure, check through my background records… but why are you looking at my record collection?
Given some of the tracks and the type of music each person listened to, could the models correctly predict what other tunes they would enjoy? A recommendation model’s accuracy was measured by seeing if the software’s suggestions overlapped with tracks a particular user had actually listened to and liked. When the team took an average of how well the recommendation systems tested, they found that their predictions were most accurate for ambient listeners, and least accurate for hard music fans.
Not only are fans of hard music less likely to listen to other genres, the songs they listen to within each genre are more likely to sound distinct from one another. Tracks that top mainstream charts do seem to sound the same. This makes it all the more difficult for recommender systems to find relevant music for hard music fans.
While the recommendation software deployed by the likes of Spotify and Pandora are secret, Lex said these algorithms will likely be based on collaborative filtering mechanism algorithms. These real-world systems will be more complex than the ones tested in this experiment, however, which is something to bear in mind.
“Our algorithms are more simplistic than what the streaming platforms use, particularly in terms of data they can exploit as naturally platform providers have complete access to users’ data,” she said.
The team hopes that their study will improve music streaming services for people who prefer their particular musical niche.
“If we think about the problem from the perspective of artists, who produce low-mainstream music, if their work is recommended more often, they get more exposure and interactions – which is crucial in this business,” Lex told us. “So, we hope that our research contributes to helping serve consumers better and to help low-mainstream artists get more exposure in music streaming platforms.” ®