Machine Learning: Dating’s Saving Grace?

Romantic Unemployment

Over the past ten years, the phrase “online dating” has undergone a social de-stigmatization. Once dubbed socially unacceptable and reserved for older “untouchables”, it is now 61% of Americans aged 18-29 and 44% of Americans 30-59 are currently using a dating site/app or have used one in the past(1). More broadly, recent studies suggest that one in five relationships are initiated online and by 2040, 70% of couples will have met via click, swipe, or double-tap(2). And yet, overwhelmingly users feel as though dating apps waste the very precious thing they sought to protect: time. As one woman poignantly put it, “there are hundreds of timewasters, losers, and just general muppets on there who have nothing better to do than mess you around.(3)” While the exploding $3bn online dating industry has attracted a multitude of players – ranging from demographic-specific silos like to quasi-intellectual elite seeking individuals on The League – many users have found the endless swiping sessions to be frustratingly fruitless (4). Nobel Laureate Alvin Roth describes dating markets as matching markets: “To work well, they have to overcome all the problems markets have to overcome.(5)” In an effort to resolve this issue of undesirable matches, companies are now turning to artificial intelligence to help improve the selection process, and consequently, the quality of the user experience. According to VentureBeat, “today’s dating sites are only as good as the data they’re given.”

Machine Learning to the Rescue? The “Relationship App” Offers A Fix

Founded in 2012, Hinge originally sourced ideal matches based on the Tinder-ized swiping approach that collated a pool of candidate based on the friends of friends of your Facebook network. In 2017, the company underwent a transformation; rebranded as the “relationship app”, the company sought to provide create more meaningful connections by showcasing possible matches in a vertical timeline format and prompting dates to like and respond directly to specific photos or prompts to match with each other(6).

Earlier this year, Hinge released the “Most Compatible feature”, which “takes into account how people act on the app (such as who and what content you previously liked). It wants to serve as a virtual matchmaker and aims to find people similar to those you previously matched with on the platform.”(7)

In order to produce these matches, Hinge uses machine learning and the training data received through a user’s interaction with the app, i.e. preferences through liking and passing choices, and then use that information to suggest a match whose preferences best align. CEO Justin McLeod describes the pairing in an interview with TechCrunch as “the best pairing we think we can find [in our user base].” (8)The app’s technology categorizes people based on who has traditionally liked them in the past. It then finds a pattern in those likes and suggests matches based on who other users also liked once they liked this specific person, a la Netflix’s recommendation feature. Both users then receive a recommendation and are given 24 hours to accept.

This process is based on the Gale-Shapley algorithm, which was designed in 1962 by mathematician and economists David Gale and Lloyd Shapley in response to the stable marriage problem, which can be summarized as followed:

“Stable Marriage Problem states that given N men and N women, where each person has ranked all members of the opposite sex in order of preference, marry the men and women together such that there are no two people of opposite sex who would both rather have each other than their current partners. If there are no such people, all the marriages are “stable”(9).

In early tests of this newly released feature, the company has found that users were “8x more likely to go on dates (as signaled by an exchange of personal phone numbers) with matches found through Most Compatible than any other Hinge recommendations.”(10)

Swipe Right: Next Steps

The company also released a user feedback feature called, “We Met”. Aimed to help better improve Hinge’s recommendation, “We Met” allows two members who were matched to confirm that they went on a date and then answer a few questions and share how that date went (11). In an effort to better improve the flow of information, the Company should incorporate more open innovation and crowdsourcing techniques in its’ product development and innovation. In addition to the “We Met” feature, the Company should allow users to link their other social networks like LinkedIn as well as their calendar and organization apps like iCal in order to inform the app on behavioral patterns. Additionally, the app can further improve their understanding of matching tendencies by allowing users to upload photos of ex-partners. By using information such as which events a user is attending or who a user has traditionally been attracted to, the app can perhaps better provide match recommendations.

While these recommendations have merit, there a few considerations that warrant hesitation, specifically: (1) how should Hinge navigate around privacy concerns and information provided on non-users and (2) how will the app deal with non-committal users (i.e. users that do not use the app frequently or provide inconsistent feedback?)


  8. See 8


Crowdsourcing improvements to America’s security: Bug bounties and open innovation in the U.S. Department of Defense


Are Machine Learning Benefits Worth Cyber-Security Risks at Chevron?

Student comments on Machine Learning: Dating’s Saving Grace?

  1. I believe the WeMet feature is a game changer. Harnessing that data reminds me of the StitchFix Model.

  2. This is an incredibly interesting application of machine learning; thank you for doing the research and sharing! This could be genius, but also I have a ton of reservations about where such technology could go. Dating definitely feels like an area to me where ‘input in’ can have a dramatic and potentially negative impact on output. I wonder if / how Hinge has thought about biases in dating preferences and what role they play in potentially re-enforcing them. I’m imagining the potential for, for example, fewer partners of a different race being presented. Are there ways to correct for this (or is Hinge correcting for this already)?

    I’m also curious about how much of the input is related to the way users interact with the platform (does it really matter that this person regular likes photos instead of quotes?) vs. other ways an algorithm could be taught to assess compatibility (for example education and work, or potentially some type of machine learning algorithm to judge attractiveness). It seems right now it’s more of the former, but if Hinge found success with the latter, would they move toward it or incorporate it into their algorithm? This seems like an interesting, potentially slippery slope to me. For example, does that reinforce some level of class-ism? Does that even matter since people are probably doing this effectively already through their own assessment process?

  3. This is really interesting, thank you for sharing! To answer your question about privacy concerns, I definitely think that is a potential issue, especially given today’s climate. However, I think they can use any data provided to them by users as a result of using the apps. The issue with privacy seems to really arise when data is sold to third parties or is used for marketing purposes.

    Part of me wonders if machine learning will truly be able to improve the quality of dating app matches. In addition to your point that the quality of the app is only as good as the quality of the people on it, I question how different real humans are from who they appear to be on their profiles. When constructing a profile and engaging in conversations on a dating app, people have access to time – time to craft responses as well as time to edit and carefully curate the photos used on a profile. This likely differs meaningfully from who most people are in real life – how will a machine be able to account for that discrepancy when creating matches?

  4. Really interesting post! I think that the promise of dating apps (a more efficient way to date) has been marred for a lot of people, as you mention, due to the paradox of choice. It is too easy to see a ton of options, and so people don’t feel the need to commit to one, or even to follow up on one. In that way, this “most compatible” feature that Hinge is launching is super interesting. Given that their is just 1 “most compatible” per person, it really makes it feel like a more serious, deeper connection right off the bat, and their is no other “most compatible” option you could alternatively choose. Therefore, I am not surprised that there is an 8x message rate. However, I wonder about the viability of machine learning to find the “most compatible” option, because I think the data set is inherently flawed. They are using a “click” as a successful interaction, but due to the nature of the app, all of the people that I clicked where people that I did not start a relationship with (or I would no longer be on the app). So, is that really a successful outcome? Therefore, I would think my “most compatible” match would be someone who I was excited to click on, but no more likely to end up having a successful relationship with.

  5. Consider this a super-like…a great, engaging article. A few thoughts that are brought to mind: what is the role of the friend in this? I thought Tinder’s use of group dates was rather innovative, and brought a more social angle. And of course, the risk with machine learning is that it relies on the training data provided, which in itself relies on personal judgement made on one’s self. Perhaps a friend could add their thoughts independently? Or is this a bridge too far for embarrassed users?

  6. I found this to be a very interesting piece on a lifestyle trend that seems to be growing exponentially. There’s a very thought-provoking episode of Netflix’s series Black Mirror titled, “Hang the DJ”, which analyzes the pros and cons of machine learning and dating apps that you may consider giving a watch.

    A few concerns that came to mind when thinking about the integration of machine learning and dating apps are the following:

    1. When identifying “undesirable” and “desirable” matches, what is the algorithm doing to promote diversity within relationships rather than simply linking people together who share relatively homogenous characteristics?
    2. Does machine learning posses the capability to match people based on attributes other than physical preferences given that most of the user input during the “matching” process derives from users making a decision based on photos of the other individual?

  7. This was such a great read, thank you for sharing! It’s crazy to see how machine learning is permeating every aspect of our lives from what we watch to who we date! It was very interesting to learn how machine learning is being used to push customized date “options” similar to what Netflix is doing with content. My fear, however, is that by doing so, a users options would be limited in terms of “playing the field” or “exploring options”, a common desire for those turning to dating apps, who often want to diversify their current option pool. With respect to matching you to people who are most like you through sharing your iCal etc., I also question whether this is the right move. Often times, people are most compatible with people who are different than them – as they say, “opposites attract”, no? Again, I fear that this may limit a user’s options.

  8. Great article and thanks for sharing! Having read into this ‘Hinge’ thing, it looks like it’s aimed at creating lasting relationships rather than more short-term ones (e.g. Tinder). This therefore makes the feedback on ‘matches’ not a particularly great feedback loop on whether it is fulfilling on that value proposition.

    Like another poster mentioned above, I think the introduction of the ‘We met’ function is a real game-changer, because now you have the company collecting data on what the interaction between the pair was like, which can then help inform how future possible matches are recommended. It would be interesting to understand what relationships exist between ‘matches’ and other on-platform activity, compare to off-platform activity that Hinge gathers from users.

    While this appears in its early stages for now, in the future the company could collect detailed information on the interaction, which could significantly increase the odds that its recommendations fulfil user expectations!

Leave a comment