The AI system that predicts a movie’s audience by analysing trailers

Share this Article:

Could an AI machine learning system be our best long term hope for original films from movie studios? Sarah investigates Merlin, and how it was used to help find an audience for Logan.

The public have pretty diverse tastes when it comes to movies. Studios don’t really know if the projects they’re investing in will appeal to viewers, particularly if they’re producing something original. After Netflix’s success using algorithms, many studios started recognising the importance of analysing data to predict the tastes of movie-goers.

One studio had been more innovative than the rest.

In 2018, 20th Century Fox (RIP) used AI to compare the trailer of X-Men spin-off Logan to several other movie trailers. If it could identify which movies were similar to Logan, Fox would then be able to predict the different audiences who would be interested in seeing it.

It wanted to find out if a superhero film with dark themes like Logan could attract a slightly different audience to the film’s expected core fan base.

Predicting the viewers of a movie in theory enables studios to have a better idea of which movies to invest in, especially if said studio is looking at challenging films that are original or cross otherwise  traditional genres.

Fox’s research centred on movie trailers because it figured – correctly – that those promos are at the heart of a film’s marketing campaign and were the key marketing tool to entice viewers to watch a film. They also show the dynamic nature of a movie, something a script can’t do. Fox thus worked with Google to create something called Merlin, an “experimental movie attendance prediction and recommendation system.”

Merlin, then, analyses a trailer frame by frame and labels the objects that appear. The whole point was to see if trailers with the same labels attracted similar audiences. In a  blog post, researchers explained that they were using a technique called machine learning, which essentially detects patterns. Could it ultimately link objects in trailers to audience behaviour?

Researchers at Fox also published a paper underlining the superiority of a system that was based on video imaging rather than text. This was a new approach. No other company researching AI and movies had really focused on trailers before. “The fact that our video-based model is able to surpass a text-based model hints [at] a new research avenue into utilizing more and richer multimedia contents to improve the movie recommendations”, it read. A bit of newspeak in there, but you get the idea.

Merlin thus scanned the Logan trailer and labelled the top ten objects as follows: tree, facial hair, car, man, vehicle, atmosphere, beard, forest and light.

Graph showing the top ten objects in the Logan trailer in descending order

But it didn’t just look at what appeared on screen, it also examined when an object appeared and how often. The temporal sequencing of objects in a trailer is very important. Knowing whether a film has long shots or intermittent short shots can give us clues as to a film’s plot and genre. A lot of long shots in a trailer typically means that the film is a drama. This is definitely true of Logan which has many long shots of Hugh Jackman, looking weary, brooding and bleeding. In one shot he just stares into a mirror.

All the data from the Logan trailer was compared to that generated from other trailers. Merlin then used this information to predict what films people who watched Logan were interested in. Fox wanted to see if these predictions could give them a more detailed breakdown of Logan’s audience. It tested the accuracy of Merlin’s predictions by comparing its data with the top 20 films Logan viewers actually went to see.

 Table showing the actual films Logan fans went to see & Merlin’s predictions

Merlin produced a list of 20 films and correctly identified the top five other movies that Logan viewers watched. It just got them in the wrong order. By predicting that Logan fans also watched films like X-Men Apocalypse, Doctor Strange and Batman V Superman: Dawn Of Justice, Merlin confirmed what researchers already knew, that Logan’s core audience was part of the superhero fan base. Granted, most of us could have guessed that fans of an X-Men movie might be into Logan too, but that notwithstanding, there are some less obvious crossovers on that list.

For it also predicted something that researchers weren’t sure about.  Merlin correctly identified that John Wick, a film about a weary action hero was in the top five films Logan viewers watched. It also predicted that audiences would be interested in action films following a rugged male hero, such as The Magnificent Seven and Terminator Genisys. Merlin ranked them too high up the list, but it was also right: those titles were in the top 20 films Logan fans watched. It appeared to demonstrate that there was an audience for Logan beyond the superhero fan base.

However, Merlin did make some mistakes. It inaccurately suggested that Logan fans would be interested in The Legend Of Tarzan and The Revenant. This was probably because forests, trees and light feature prominently in The Legend Of Tarzan. In The Revenant, it was the prominent beards, facial hair and forests that made Merlin mistakenly suggest that it was a film Logan’s audience would be interested in. It also missed that unconventional superhero films such as Ant-Man and Deadpool had the same audience. But with time and improvements in  AI, it’s hoped that more and more mistakes can be ironed out.

After all, Merlin is programmed not to be static. It’s an ongoing project whose data is updated weekly so that it can account for new trailers coming out. This means that it can analyse data about the newest releases. After the Logan trial, researchers created reports and prototype marketing campaigns from Merlin’s data. They also tried to improve the software by incorporating DVD and streaming sales and rentals into the model.

In theory, going forward, Merlin could be used to broaden a film’s audience behind the expected fan base. Its predictive software could change how a film is marketed or even influence whether or not a film is made. Breaking down an audience into micro segments could lead to studios taking more chances. It may yet be our best chance for ongoing original films from major movie studios, an odd indicator of the times we live in. That said, for Fox, it didn’t quite pan out longer term. But that’s a whole other story…

Thank you for visiting! If you’d like to support our attempts to make a non-clickbaity movie website:

Follow Film Stories on Twitter here, and on Facebook here.

Buy our Film Stories and Film Stories Junior print magazines here.

Become a Patron here.

See one of our live shows, details here.

Share this Article:

More like this