The 'magic formula' still eludes us
Phone phreakers, scammy apps, Tiktok villains and the surprisingly intractable "hit song science" problem
Hi friends. Today is January 28, 2022.
And for much of the past week, major music charts have been absolutely dominated by the surprise Disney hit "We Don’t Talk About Bruno.”
“Bruno” doesn’t precisely scream “Billboard Hot 100.” It’s a minor earworm off Lin-Manuel Miranda’s Encanto soundtrack. For three-and-a-half minutes, an ensemble cast trades anecdotes about the movie’s creepy uncle character over a slinky, salsa-y, off-beat rhythm. Its lyrics make absolutely no sense unless you’ve seen the movie (unlike, say, “How Far I’ll Go,” the big hit from Moana). And listeners can’t exactly generalize its themes to their own lives (unlike “Let It Go,” which I have personally blasted to exorcise the memory of many a bad work/family/romantic encounter).
This is often how hit songs work, though: It’s hard to predict what will stick. Until very recently, in fact, even algorithmic models trained on tens of thousands of songs couldn’t forecast a new track’s success — a conundrum known, in computer science, as the “hit song science” problem. Even now, in the age of algorithmic market predictions and early prison releases and God knows what else, these musical models aren’t nearly as accurate as you might expect.
So — how do you predict a hit? Early challengers tried to tackle the problem by focusing on songs’ features: things you could easily identify and tag, like the genre or tempo or topic of the song. The first model of this type, published in 2005, claimed it could identify common threads between hits in different genres. (That was largely debunked.) Another attempt at hit song science, conducted three years later, found no direct link between song popularity and any particular feature. A third go, in 2011, also failed to predict the popularity of YouTube music videos.
Some computer scientists began to wonder if popularity was just too messy, too human, for algorithms to grasp. Writing for an academic publication in 2011, Francois Pachet — then a researcher at the SONY Computer Science Laboratory in Paris — argued that computer models would struggle to account for variables even music psychologists didn’t understand.
After all, people don’t just like a song because of its lyrics or its proportion of percussive to harmonic sounds: They’re also influenced by other people. They gravitate toward or away from the songs they hear most often, according to where they were or how they felt when they heard them. And they encounter many songs not through personal choice, but through third-party radio DJs or Spotify recommendations. How could even a really good AI predict that “Bruno” would become a breakout TikTok trend?
But at the same time computer scientists were pondering these questions, consumers were changing how and where they listened to music, giving researchers access to an abundance of new information they hadn’t previously had. While the earliest hit song science model used only 1,700 songs to train, researchers were running machine-learning algorithms on tens of thousands of Spotify songs by the mid-to-late 2010s.
In 2019, two undergrads in California claimed their model — based on a data set of 1.8 million songs, and built using "the entire fleet of computers available to [the] University of San Francisco’s Computer Science Department" — could predict Billboard hits with 88% success. Another service, founded by a senior researcher at the Finnish Centre for Interdisciplinary Music Research, also claims to accurately identify high-potential songs, and shows some of its work in a weekly forecast.
But the “magic formula” for hit prediction still eludes us, two data science professors wrote just over a year ago. No one has achieved perfect accuracy in their predictions, and the field hasn’t congealed around one particular model.
That might be a good thing, given that the ability to predict the popularity of a piece of music could change which songs (and books and movies and TV series and fashion lines and art exhibits) get produced, at all. Would a hit-song algorithm have green-lit the creepy-uncle Latin-pop song…? Dunno, but it’s hard to imagine.
If you read anything this weekend
“Searching for Susy Thunder,” by Claire L. Evans in The Verge. This is just a wild, perfect profile — of the type that persuaded me to study magazine writing in college! — on the ex-phone phreaker/groupie/con artist Susan Headley, who became the greatest female hacker of the 1980s before going underground. Among her many accomplishments: nearly taking the LA phone system offline; sleeping with all four Beatles … you honestly couldn’t make this stuff up. It is wonderful.
“The Internet is Failing Moms to Be,” by Nina Jankowicz in Wired. Nina Jankcowicz has advised U.S. and foreign governments on political disinformation. So when she says pregnancy apps rank among the worst pits of disinfo she’s seen, that seems … scary! And significant!
“Nothing Sacred: These Apps Reserve The Right To Sell Your Prayers,” by Emily Baker-White in Buzzfeed. Speaking of scammy apps, Buzzfeed published a double-header last week. First: the aforementioned, on the outrageous data practices of religion apps. Second: this one, on the (now well-documented!!) fact that Noom is really, truly just a diet.
“Silicon Valley’s Blind Spot,” by Josh Gabert-Doyon in The Baffler. Remember Blind? Me neither, really! The app, an anonymous social network for tech workers, enjoyed a short-lived bout of mid-’10s popularity. Amidst the “Great Resignation,” however, it’s gained new relevance — if only as a symbol of how unreformable the tech industry is.
“The Only Hat You’ll Ever Need,” by Dana Bell on Medium. I can’t remember the last time I shared fiction in this space, but … hoo boy, trust me on this one.
👉 ICYMI: The most-clicked link from last week’s newsletter was this story on amateur Peloton porn. I see you!!
Postscripts
Crypto-simps. Meta girlfriends. Divorce registries. World Wide Wordle and hyperpop, explained. What happened to Peloton and Netflix. The first “nonconsensual TikTok villain of 2022.” (This is, so help me God, the first and last thing I’ll link on West Elm Caleb.) Why electric cars haven’t caught on in the US. Cracking a $2 million crypto wallet. Kinda feel like, if you can’t share a password, your relationship might not be … that solid?
An appreciation of @_restaurant_bot. Easily the best “house” on Zillow. Blockedchain, “trauma” and Internet nutrition labels. @OKWildlifeDept is the new Monterey Bay Aquarium. Linktree is the new personal site. The rise of crypto mayors, TikTok doctors and — idk, maybe concerningly? — AI that builds AI. Inside the Black Manosphere and the metaverse real estate boom. The fledgling industry that helps influencers get paid. Last but not least: a profound little essay on how our relationship with photographs has changed.
That’s it for this week! Until the next one. Warmest virtual regards.
— Caitlin