Recently, I’ve read not one but two articles on using algorithms to predict the success of movies. In both cases, the analysts start with a script (one team uses human analysts, the other a computer), then analyze it for key elements, plot twists, setting, characters, and so on. Then they use what appears to be regression analysis to determine the degree of commonality the script has with other hits.
And it works. Hits show up. Misses are clearly seen.
This isn’t entirely new, of course. There have been a couple of attempts to do this with music, for example. Over at Platinum Blue they use algorithmic analysis of music in order to tell which of the sixty types of hit songs a given piece might be. They can tell you that “sorry, you’re somewhat off on the bass track and only 65% likely to be a hit.” They can’t tell you how to fix it, though.
The case has been made that this doesn’t lead to gross uniformity in the arts. After all, currently publishers and funders of all sorts routinely fail to take a flyer on unproven talent or stuff that seems a bit off the beaten track. But underneath the unproven aspects and the unfamiliar elements might be the skeleton of something completely familiar, something designed, inadvertantly or not, to tap into something deep in the human brain.
Of course, the recipe here is not for artistic success; it’s for commercial success. Those sixty hit song archetypes go through fads, they fade in and out of popularity (Beethoven and Mozart are apparently solidly in one of the clusters). Obviously, some people have this algorithm in their heads: Lennon & McCartney each seem to have had half of it, and people like Billy Joel who manage hits spanning 30 years must have a pretty good knack for it, and legendary producers like Clive Davis maybe even have some of it written down. It’s not about “good” but about “popular.”
This past spring, for instance, he analyzed “Crazy,” by Gnarls Barkley. The computer calculated, first of all, the song’s Hit Grade—that is, how close it was to the center of any of those sixty hit clusters. Its Hit Grade was 755, on a scale where anything above 700 is exceptional. The computer also found that “Crazy” belonged to the same hit cluster as Dido’s “Thank You,” James Blunt’s “You’re Beautiful,” and Ashanti’s “Baby,” as well as older hits like “Let Me Be There,” by Olivia Newton-John, and “One Sweet Day,” by Mariah Carey, so that listeners who liked any of those songs would probably like “Crazy,” too. Finally, the computer gave “Crazy” a Periodicity Grade—which refers to the fact that, at any given time, only twelve to fifteen hit clusters are “active,” because from month to month the particular mathematical patterns that excite music listeners will shift around. “Crazy” ’s periodicity score was 658—which suggested a very good fit with current tastes. The data said, in other words, that “Crazy” was almost certainly going to be huge—and, sure enough, it was.
In The Long Tail, Chris Anderson makes the case that hits are actually less emotionally satisfying that media that targets an individual’s particular likes and dislikes. They push a lot of buttons, but not very well.
You can think of our tastes as being like a box of springs. We all have the same set of springs, but they are tuned differently, some tightly coiled and some not. When a spring is depressed all the way, whatever is resting on it is “hitting our buttons.” A hit is something that lays across most of the springs with some moderate weight. It effectively presses all the buttons, just not very well. A tightly focused niche title aims for just a few springs, and presses down really hard; if you have loose springs there, you’re gonna love it. Everyone else “won’t get it.”
But again, we’re not talking about artistic merit. You see, the springs are for things like harmonic overtones, excitation in specific frequency bands, that sort of thing. They are about mechanics, about execution, and when you reduce the world to clockwork, sometimes the art might fall out. (Yes, I’ve mentioned this before.
The only reason this can be done with movies and with music is because there is enough of an established atomic structure to these media that you can do analysis on it. Even with something that seems as subjective as a story, there’s enough elements that can be isolated that a neural network can chew on them. (Check out the revisions to The Interpreter in Gladwell’s article to see how it works).
Games, however, lack this. We instead have something more akin to the disastrous attempts to make the “ideal painting,” that were done purely subjectively, via opinion polls. And the thing is that people don’t actually know their own buttons very well.
Sixty-seven percent of respondents liked a painting that was large, but not too large — about the size of a dishwasher (options ranged from ‘’paperback book’’ to ‘’full wall’’). A whopping 88 percent favored a landscape, optimally featuring water, a taste echoed by the majority color preferences, blue being No. 1 and green No. 2. Respondents also inclined toward realistic treatment, visible brushstrokes, blended colors, soft curves. They liked the idea of wild animals appearing, as well as people — famous or not — fully clothed and at leisure.
Can games get there? Yes, I think they can, but it’ll take analysis at a deeper level than what we have currently, followed by an exhaustive historical survey of games to build the database and clusters.
And then we’ll get to have the debate about whether we should be doing it at all.
In the meantime, Platinum Blue offers a consumer service at $10 a song, and I am inclined to give it a try.