Can a computer learn when to swipe right on Tinder?

A researcher in Canada says a computer achieved 68 percent accuracy

150914 tinder de vries

An artificial intelligence programme to improve Tinder suggestions has been developed by Harm de Vries, a post-doctoral researcher at the University of Montreal who was sick of swiping left.

Credit: Université de Montréal

When it comes to the complex laws of physical attraction, can a computer learn your preferences on Tinder?

It sounds like a tall order. Ask someone to describe their "type" and you'll often get a list of a few simplistic attributes. But dig a little deeper, and what attracts one person to another reveals itself to be much harder to pin down. It's laws are so mysterious they're often explained as "je ne sais quoi."

But after hours of swiping profiles on the Tinder dating app, Harm de Vries, a post-doctoral researcher at the Université de Montréal, wondered if a computer could help him a little with his search.

"Tinder kept giving me pictures of girls I wasn't attracted to," he said in a phone interview. "So I wondered if I could use deep learning."

If a computer were given enough examples of when he swiped right (like) or left (dislike), it should be able to figure out his preferences, he thought.

De Vries began by scraping 10,000 photos from Tinder and giving the computer his opinion on 8,000 of them. The computer tried to determine his likes or dislikes for the remaining images but managed only a 55 percent success rate -- only slightly better than flipping a coin.

De Vries decided the computer needed more images to choose from, so he pulled almost 500,000 photos from OK Cupid, another popular dating site. The computer's first job was to determine men from women -- something it achieved with 93 percent accuracy. De Vries himself managed 95 percent accuracy (most of the 5 percent he missed can be attributed to people uploading pictures of pets or food).

He then fed the additional images into his original program and tested it again. With a greater number of images to work with, the computer achieved a 68 percent success rate -- in other words, de Vries agreed with two out of every three choices the computer made for him.

"It's a pretty good start," he said, "but I think it can be improved a lot."

Still, he acknowledges it's a tall order. "One of my friends who collaborated with me got to learn my preferences and he managed 76 percent accuracy, so even for humans it's pretty hard."

Based on his results, de Vries thinks services like Twitter could benefit from machine learning, but he also knows that attraction is an inexact science.

"Your brain might look over a picture and even though you can't describe why, you can still say whether you like it or not," he said.

So has his research improved his dating life? De Vries said he's been on a lot of dates because he's been doing so much swiping, and the publicity generated by the study has helped him as well.