Thursday, August 06, 2020

Thoughts on Dating Apps

I have been thinking about the  problem that computer dating is designed to solve. I can see four different approaches, three that exist in both realspace and online, one that could exist online but I am not sure does.

1. Examine the characteristics of many people, looking for someone you believe would work as a partner.

Subscribers to OKCupid answer questions about themselves, if they wish many questions. They look at the answers of other subscribers to find someone who looks like a good fit. It is an approach that people routinely use in realspace, considering people they encounter of the appropriate age, sex, and marital status as possible partners — except in a society such as Saudi Arabia where contact between single men and women is very strictly limited, leaving it to a man's mother and sister to find him a suitable bride.

The great advantage of computer dating is that it lets you search a much larger number of potential partners than you would encounter in realspace. On the other hand, the information you are using is provided by the potential partner, designed in part to make him or her appear a good match, so may be less reliable than what you could learn by direct observation.

2. Search a set of people already selected for you to have characteristics that make success more likely.

OKCupid allows you to filter for people within a given geographical area, by age, gender and some other characteristics. Dating apps as a group carry this farther, specializing in Christians who want to marry Christians, Jews who want to marry Jews, individuals looking for casual sex, individuals interested in marriage and uninterested in casual sex, and  other categories of potential subscribers.

On the other hand, computer dating does not use the information implicit in who you associate with in real life, mostly people who have a good deal in common with you. A student at Harvard or Vasser is associating with other people who have passed the same filter, making a fit much more likely than with random pairing — this may be one of the major functions of institutions for higher education. Joining a political movement, becoming involved in a hobby, entering a profession, all put you in contact with a set of potential partners who share with you characteristics that might lead to a good fit.

3. Use an intermediary, someone looking at individuals and trying to match them up.

In the case of computer dating, the program can be the intermediary, showing you not a random collection of individuals but individuals that, in its judgement, fit what you are looking for and are looking for what you fit. The obvious way of doing this is based on the subscriber's account of his preferences for a partner but it also could be, perhaps sometimes is, based on someone else's theory of what characteristics go well together.

The realspace version in more traditional societies is the shadchen, the marriage broker. In a society such as modern-day America, it is done more informally. A colleague's wife suggests that there are a lot of nice girls at folk dancing — I go, despite not being a dancer, and meet the woman I have now been married to for almost forty years. I ask the wife of my married son, a perceptive woman who knows lots of people, if she can suggest anyone my single son or single daughter would get along with.

4. Use feedback on the results of attempted pairings to discover what characteristics predict successful ones.

The first approach depends on individuals knowing what sort of person will work for them as a partner. The third depends either on their knowing that or on someone else figuring it out for them. There might be a better way.

I am imagining a dating app one of whose terms is that, after your first date with someone met online, you provide the app with feedback, report whether things went well or badly, with further reports after any later dates and a final report after you either give up on each other or decide on a long-term relationship. The program uses that information to figure out what characteristics, defined by the information provided to them by each subscriber — not "is X," which the program does not know, but "says he is X" — predict a successful match. The fact that I had a successful or unsuccessful date with someone with a given set of characteristics not only predicts what other people I will have successful dates with, it  predicts the same thing for other people who share my characteristics. Thus, over time, the program could provide experimental evidence on successful pairing based on self-reported evidence, making possible a better version of the previous approach.

Have any dating apps actually done this? If so, with what results?

I originally came up with the idea in the context of matching readers with books, inspired by a talk I heard on multi-dimensional voting theory, an approach to matching the positions of voters with the positions of politicians in order to predict electoral outcomes.

10 comments:

Mitch said...

Isn't this Netflix's Cinematch for dating? (Their movie recommendation engine which was already excellent and for which they offered a $1M prize for an algorithm which could improve it by 10% before abandoning that whole effort in favor of whatever execrable system they now use)?

Albert said...

Even if the fourth approach worked, in today's climate it would never be allowed. There would be too many factors such as race, physical appearance, or wealth that would be predictive but considered unacceptable biases. Look at the furor over GPT-3.

Anonymous said...

Interesting moral problems for dating apps that can detect people's true preferences:

Suppose there is a woman who has a hidden preference for domestic abuse. It surely causes her much mental torment, but that is the kind of men she is most attracted to. Should the algorithm match her with a man who is prone to engage in domestic abuse?

What about if we suppose that the app could have an effect on human evolution? Say that two people that need glasses due to bad eyesight make the best pairing in the view of the app. Shouldn't then the app make it so that people with poor sight have the most children, so it can maximize the best pairings in the long run? I guess that would be nice for the app, but humanity as a whole would suffer due to poor eyesight. Should you allow this kind of development to occur?

Actually, the best business strategy for a dating app would be to maximize the amount of users it has. The most users make the most money, no? Then it would be rational for the app to start breeding humans that pass down the app usage tradition to their children. The app would then become kind of a human breeding program, that maximizes the amount users by creating more humans. I find that this could be the best long run business strategy.

NZ said...

@Anonymous: I wrote a short story along those lines: nzfictionsandbox.wordpress.com/2019/09/23/a-prayer-from-2332/

Markus Rieder said...

Thanks for outlining the different approaches so clearly, David. In fact, two colleagues and I had the idea to leverage the user experience (and even the experience of non-users, but successfully paired couples) to develop a matching algorithm. As an input to this algorithm served a set of (personality) characteristics, and the output was a simple matching score predicting the success chance of a certain pair.
The results were astonishingly accurate from a statistical point of view, so the model worked quite well. The story behind the formula, in short, is also well-known: Pair with an equal, since all characteristics that differ between you and your opponent increase the risk of separation dramatically. But the economic success of our dating app was meager. Neither any of the big dating platforms is interested in a good matching rate, nor did the users believe that age and education status is relevant but your favorite color or your zodiac sign is not.
Still, if data sets increase (in both width and length), making predictions even better, and people get more inclined to trust automatic judgments, we might pick up this work again some day...

Seth said...

Hi David,

a few things based on my experience with OkCupid:

> after your first date with someone met online, you provide the app with feedback, report whether things went well or badly, with further reports after any later dates and a final report after you either give up on each other or decide on a long-term relationship.

Today, if you disable your OkCupid account, you are prompted to give a reason. One of the options is "I met someone on OkCupid." If you select that option, it will then ask you if it's one of the people you've exchanged messages with lately, along with their pictures.

So perhaps they are using that data as you predict. The App also knows when you sign on again, i.e. when you are single again.

My 2C: the first main advantage of more dating activity moving online is that it's physically safe to screen out bad/unwelcome messages. You just delete the message and block the person and that's it. If you romantically reject a man in person, there's a small but real chance they're another Elliot Rodgers in the making.

This doesn't eliminate the threat of a man turning violent if you reject him in person, but on the margin, even for those men, OkCupid/Tinder/etc. probably does push some of their behavior online, where it's easier to deal with.

The flip side of that benefit is reducing false-positive date requests (asking someone out when they don't want you to). If you match with someone online, you know that they are:

A) single
B) potentially interested in you
C) Open to being asked out on a date.

You don't know these to be true for someone you meet in your book club (though the matchmaker/introductions-through-family methods also figure these out on your behalf). You can suss it out, but there's a chance you'll get it wrong, as either a false positive or a false negative (*not* asking someone out who wants you to).

For men today, the costs of a false positive are perceived to be much higher than they used to be.

Assuming you can write a pleasant, non-harassing message to someone you've matched with on OkCupid, you've greatly reduced the probability of false positives. This is attractive to many people.

The main thing lost from online dating, IMO, is you don't have whisper networks/gossip. As they say, on the internet, nobody knows you're a dog, so if a man's college classmates all think he's a creep, if he can write a nice message, you don't know that.

You lose information, and the whole thing can feel a little scripted, but all parties gain a lot in safety and clarity.

David Friedman said...

@Markus:

What was your definition of "success?" For someone looking casual sex, "success" means sex. For someone looking for marriage, it means a long term stable marriage. For a lot of people, it probably means "this is someone I would date again," on the assumption that the finer filtering is going to be done in realspace.

Markus Rieder said...

@David:
For the dating app we developed, we built upon a large data base that collected several personal traits and characteristics of individuals starting at the point in time when they began some kind of relationship. This person-specific information served as the independent variables in our model. The couples have been observed over several years and tracked with a periodic questionnaire then. For the dependent variable, we defined "success" as the duration of the relationship.
So, yes, couples might have different goals when matching on a dating platform. However, we assumed that a shorter relationship indicates a poorer match, where a longer relationship would hint to a better fit. Maximizing the relationship duration still seems to be a good idea for customers (even when looking "only" for sex!), but maybe not so for the platform providers who'd rather prefer maximizing the dating app retention time...

Joe said...

Is the dating app still around?

Mr Bitcoin said...


In Freakonomics, which was written in 2007?, Levitt says that online dating preferences are clear once you filter out the noise.

Men prioritize women are young and thin.

Women judge men based on height and income.

A picture is important. I think it's more important for women but I can't recall.