Swiping Right on AI: How ML Models Can Learn to Play Cupid from Your Data

Searching for the right match on dating apps can be overwhelming because of the huge number of options and limited time. That’s why many apps use Machine Learning Algorithms to help filter through the choices. Since almost everyone is familiar with dating apps, I hope that this piece is very accessible & easy to understand. I’m well aware that dating apps use far more sophisticated approaches, alot of which are kept under wraps. However, the main goal here isn’t to crack the Hinge Algo but rather to create an accessible example of how AI/ML models work & are applied.

Imagine being in a crowded room where everyone is looking for a date. You chat with a few people about your interests and soon discover others with similar preferences. Through these connections, you’re introduced to candidates that match your taste, leveraging the collective insights of your newfound acquaintances. This process is at the heart of collaborative filtering: finding matches based on the preferences of like-minded users.

Say you are the biker, you know the runner is a similar person to you because they exercise and eat same foods. Since they also drink energy drinks its likely you might enjoy energy drinks because you share show many interests. This is the general logic behind collaborative Filtering

There are AI/ML models that apply the general principles of Collaborative Filtering. Collaborative Filtering is only one of the approaches used by dating apps but it provides a great introduction to AI/ML

How it works:

Through the app, millions of data points have been collected on you & others as to what features are most predictive of right or left swipe on a profile.

It compares your profile and interactions with those of other users to find users whose likes and dislikes are similar to yours based on the aforementioned data collected

Based on these similarities, the system generates a list of profiles you haven’t interacted with yet but are likely to be interested in, based on the patterns identified.

Explanation using data: This is a high-level approach but hopefully it makes it more real for you as to how these models work.

Here is the dataset where you have 5 users who interacted with some combination of 5 profiles. What’s been captured is how they swiped on them & what rating they gave them, and how much they have used the app.

Say the goal is to provide recommendations to User 1, step #1 is to identify users that are similar to User1 through the application of complex math.

What is output is a proximity score, so we can see Users 2 & 4 are the top users most similar to User #1.

• We sum up the ratings for each profile from similar users, weighted by their similarity score to User1.
• We only consider profiles that User1 hasn’t rated yet which we know to be C&D

The model recommends that Profile D is the next profile shown to User 1 because it believes based on the scores provided by Users 2 & 4, there is a high likelihood of Profile D also being appealing to User 1. How User 1 reacts to Profiles D is captured & used as training data the next time the model runs.

As mentioned earlier, this is one of many approaches used by dating apps all of which come together to provide value to users. Hopefully, you find this accessible so we can explore other algorithms used to demystify AI/ML further.