🔥 Tinder : what chess and matches have in common

My motivation in this article is to show how dating apps apply basic matchmaking algorithm inside their apps. I chose Tinder as main example, because the app is well-known and the algorithm has been studied many times.

What made Tinder successful for years is a well-known and simplistic algorithm, let's dive in !

📈 Tinder before 2019

We are 2 years ago (2018). Here is what Tinder was offering back then.

My name is Jérémy, I am 25. I live in Montréal. I enjoy going out, playing video games, watching TV shows. I would love to go get a drink someday in order to get to know each other.

Screenshot_20201202_152019.jpg

That's how Tinder will display my profile to other users. But how Tinder can match me with someone with the same interests ? Maybe match my description with other's descriptions ? Nah, it is not possible.

Tinder will use few parameters to match me with other people :

  • my preferences settings (age, gender, geographic area)
  • my Elo score

♟️ What the heck is Elo ?

ELO score comes from board games ranking such as chess. It is a method that computes a rating for each player based on his skills.

Let's take an example you may know. In the Queen's Gambit TV show (go watch this if you didn't), the main character Beth is playing in a chess tournament.

Beth is a chess prodigy, but at this this point she is only starting her career. A young yet high-rated player, the famous Benny Watts, shows up and talks with Beth. She asks him if he will be playing.

Benny Watts explains that he is not playing in this event because it could only hurt him to play too many tournaments. Knowing that his rating is high compared to others players, why would he say that ?

benny watts.png

Let's say Benny has 2000 Elo score. Beth wants to play against him, but since she is still a newcomer on chess competitions she has 1500 Elo score. Since Benny has way more Elo score, here is what would happen after the game:

  • Benny wins over Beth : he would win 5 points and Beth lose the same amount.
  • Benny loses over Beth : he would lose 27 points and Beth wins the same amount.

That makes sense, Beth is not supposed to win against Benny based on ratings. Benny refuses to play because if he wins he gets almost nothing, but losing is a very bad news. That's a low reward/high loss situation for Benny, and he is right to refuse, even more because he knows that Beth is way better than her rating could reflect.

benny smart.jpg

So yeah, Benny seems likes a smart guy. He figured out Elo score pretty well.

📊 Tinder's Elo system

Okay, Tinder is using Elo score. It is a proven fact, they confirmed this hypothesis with statements few years ago.

But Tinder is not chess, how would you implement Elo score there ?

First of all, the metric to compute this score, often mentioned as desirability score, is the like.

When Tinder shows up a profile, you can either swipe right to like someone, or swipe left to ignore someone. Your Elo score will be impacted by the people that likes you.

swipe.png

A match is when 2 people swipes right (= likes) each other, and does not seem to impact Elo score.

Let's say your Tinder Elo is between 0 and 10. You would think your starting Elo is 5 ? Nope, it would be 6 or 7, because Tinder gives new accounts an hidden Elo boost to engage user quicker. It means that during few days, your Elo score will be higher than it should, and after that period it will go back to normal. Now you start using the app.

What happens next ?

You start swiping, and you will mostly see people around your Elo. It is not a closed range, so you will still be able to see profiles way below and way above your score, but the likeliness decreases when difference between scores increases.

Your chances of seeing a profile of a certain Elo score would be something like normal distribution.

gaussiantinder.png

By swiping right people (= like them), you might modify their Elo. Just like in chess, if you like someone with lower score, they will gain few points of Elo. But if you like someone with higher score, they will gain very few or zero score (hard to tell actual values).

As a result, your desirability score is based on the likes you get from higher score than you. By doing that, Tinder states that they think people with same "desirability" are most likely to date each other. I let you make your own opinion about that !

What might lower your Elo ?

If you make the maths, we haven't described any way to lose Elo yet.

At first, I thought that if a person with lower Elo likes you, it decreases your score. After some research, it is not a valid statement, and it would hurt Tinder success a lot if implemented that way.

Take into consideration that the following statements are only true if you are not paying for the app. Paid plans have undisclosed advantages that may affect the way Elo works for those users.

The ways of losing Elo are completely driven by Tinder's will. There is no official statement on this subject, but a lot of users found interesting mechanisms.

Apparently, you might lose Elo by :

  • swiping left too much (Tinder does not want to show profile that never matches)
  • swiping right too much (Tinder does not want robots)
  • being inactive for too long (Tinder does not want ghost profiles)

Those are the main reasons you might lose Elo based on a lot of users experiences.

If you are not following Tinder guidelines, you could get shadow-banned. It means that you will still have access to the app, but your profile wont show up to other users. Tinder wont notify you about this shadow-ban. So think twice before trying to cheat on the algorithm 😉.

🤖 Tinder transition to a more complex algorithm

In March 2019, Tinder stated that they were moving out desirability score as their only metric. They added new ways to match people together, in order to improve likeliness to have a successful date.

Yet, they kept the Elo score alive, and it is still involved in current algorithm, difficult to know at which level.

Tinder does not give its secret receipt to public, but here is what we know so far. The new algorithm is a combination of your preferences (the ones you put as settings), desirability score (less prominent than before), and a mysterious analysis of people you swipe.

The new factor is the last one mentioned, it seems that Tinder is analyzing with Machine Learning techniques the pictures of the people you swipe, in order to get some of your preferences. Tinder stated very strongly that they were not matching any sort of explicit preferences such as ethnicity, religion, kids, but since those new preferences are AI-driven we do not really know how it works.

A naive take on this would for the algorithm be to say: "A has a cat on his picture, B swipes right, let's increase the B preference's for cat". I let you make the pessimistic example by yourself. We know Machine Learning cant be human-readable like the example in my previous statement, so it is hard to figure out what actually happens.

My opinion is that this famous desirability score is still very important on the algorithm, because (sadly ?) it worked well in Tinder, that reached a gigantic amount of users using that.

My intent is not to encourage or discourage you to use Tinder, it is totally ok nowadays to use such app if you like.

My intent was really to display the hidden layer of that kind of app, and put it in perspective with standard matchmaking algorithms. A lot is yet to be discovered, and during my research I found a fair amount of people willing to know more about dating algorithm, each of them with an objective of their own 😅

I am interested in your opinion about this algorithm, there is a lot to say on this topic, so do not hesitate to jump on comments ⬇️

Thanks for reading !

No Comments Yet