HIT songs are big business, so there is an incentive for composers to try to tease out those ingredients that might increase their chances of success. This, however, is hard. Songs are complex mixtures of features. How to analyse them is not obvious and is made more difficult still by the fact that what is popular changes over time. But Natalia Komarova, a mathematician at the University of California, Irvine, thinks she has cracked the problem. As she writes in Royal Society Open Science this week, her computer analysis suggests that the songs currently preferred by consumers are danceable, party-like numbers. Unfortunately, those actually writing songs prefer something else.
Dr Komarova and her colleagues collected information on music released in Britain between 1985 and 2015. They looked in public repositories of music “metadata” that are used by music lovers and are often tapped into by academics. They compared what they found in these repositories with what had made it into the charts.
Metadata are information about the nature of a song that can give listeners an idea of what that song is like before they hear it. The repositories presented Dr Komarova and her team with more than 500,000 songs that had been tagged by algorithms which had been trained to detect numerous musical features. The tags included a dozen binary variables (dark or bright timbre; can or cannot be danced to; vocal or instrumental; sung by a man or a woman; and so on). The team fed all of this information into a computer and compared the features of songs that had made it into the charts (roughly 4% of those in the repositories) with those of songs that had not.
Overall, the team’s results suggested that songs tagged as happy and bright have become rarer during the past 30 years; the opposites have therefore appeared with greater frequency. That was not, however, reflected in what made it into the charts. Chart successes were happier and brighter (though also less relaxed), than the average songs released during the same year. Chart toppers were also more likely than average songs to have been performed by women. All this is important information for executives of music companies.
Dr Komarova used these results to train her computer to try to predict whether a randomly presented song was likely to have been a hit in a given year. The machine correctly predicted success 75% of the time, compared with the 4% rate that guessing success at random from the music database would yield—something else music executives might pay attention to.
Content is not everything. As might be expected, circumstances—particularly any fame already attaching to a recording artist or artists—had an effect, too. But not a huge one. Adding in information about who was performing a song increased the accuracy of prediction to 85%. That suggests that musical fame is actually attached to talent, rather than to hype. And this, perhaps, is a third lesson for an industry that some believe is not wedded to talent enough.