They are located in an old building at the Schuitendiep, but any hint of being timeworn end at the threshold. The modern space houses three guys, each intently staring at his computer screen. Lego robots, small computers, and 3D printers are scattered around the building. A beamer lying on the ground next to a table seems a bit lost. HitWizard is nowhere to be seen.
Seated at the long conference table by the window, Ard Boer, who studied arts, culture & media in the distant past but never graduated, opens up his laptop. ‘This is what HitWizard looks like.’ On a white screen, we see an input field over a generate button. At first glance, it does not look like much. But Boer and his colleague, Zakarias Nordfält-Laws, worked on the underlying algorithm for three months.
Previously, Boer came up with ‘Radio Radar’, a programme that maps out when and how often tracks were aired on various Dutch radio stations. ‘That formed the basis for HitWizard’, Boer explains. ‘Zakarias knows all about artificial intelligence, and I had access to airplay data because of Radio Radar. Together we thought: how can we use that data to teach a computer when something will be a hit or not?’
Neural network
The airplay data supplied by Radio Radar are leading in the algorithm. But missing from the equation are musical characteristics. Using software, the developers requested those characteristics from Spotify. ‘We work by the grace of the Spotify’, the Boer says, laughing. The streaming service collects musical data, which is often based on its own algorithms. ‘For every track, we know how long the song is, its tempo and key, but also how energetic a song is’, Boer explains. But data alone is not enough to complete a computer algorithm. It also involves thinking, of the human kind.
‘For example, you have to determine which mechanisms to use’, Boer says. The developers decided upon a neural network, which means there are different values that influence and inform each other. Next, the computer trains itself to find the underlying connections between the parameters and to link these to a control value, which defines when something constitutes a hit. In HitWizard’s case, something is a hit when it is on the list of the 200 most played tracks on Spotify.
How do you teach a computer to learn?
Boer explains the principle of an algorithm: ‘A good example to use in explaining how an algorithm works is that of cat and dog pictures: if you show a computer 1,000 pictures of cats and 1,000 pictures of dogs, there’s a good chance that the computer will recognise what’s on the 1,001st picture. That is because in that time, the computer has learned the different characteristics of the animals. The fact that an animal has four legs is irrelevant in this case and will be excluded. But characteristics such as fur, nose size, or the distance between mouth and eyes are. We applied the same principle to HitWizard.’
Beta
HitWizard is still running in beta mode. The algorithm’s predictions about which songs will become hits are 60 per cent accurate. HitWizard’s predictions about when a song will not become a hit have been 99.3 per cent accurate. ‘That’s the best result we’ve had so far’, Boer says. In the future, he hopes to improve upon the algorithm. But the guys are not entirely sure which characteristics are of importance in predicting a song’s hit potential. ‘Much more interesting is the fact that we discovered that SlamFM and Radio538, for example, are really good at predicting hit potential. Another radio station you might expect to be quite good at that, 3FM, actually isn’t.’
Boer and Nordfält-Laws hope to find a big market for their product. ‘We’re just trying to see if a computer can predict whether something will be a hit or not as well as humans can. After all, that is what the music industry is all about. The only thing that judgement is currently based on is gut feeling. And whether people like it or not, machine learning is becoming more prominent in our lives.’
New things
If HitWizard is a success, will that lead to the downfall of music? Boer answers the question with a broad grin. ‘A friend on Twitter asked that same question. But I don’t think it will be possible even in a hundred years to have a computer consider all the parameters needed to determine whether something will become a hit or not. That’s because for a lot of parameters, we simply don’t know if they influence anything at all. And besides, people will always have this need for new, fresh things.’