The Data Science of Dating Dan Kras https://desci.nyc/transcripts/the-data-science-of-dating Automatically generated transcript. May contain errors. [0:07] you can hear me okay cool um thank you guys all for coming I appreciate uh uh everyone coming out here tonight this is an amazing turnout um hopefully uh a lot of you find uh like learn a lot about dating and have met some really cool people here um and it's been pretty cold out so I appreciate everyone coming all right we'll just jump right in since there's like so much to do and it's like kind of a crazy crowd um so if you don't if you want to find out about future events um you can go to dnyc on Twitter so you can take out your phone and follow it you guys got to take out your phone come on you can do it has any here know what decentralized science is are [0:48] you guys on Twitter and like have you seen it popping off recently no okay yes okay good somebody is that an actual yeah I come to these events okay yes okay good we don't want authentic yeah oh what is it so we'll talk about it in a second um but decentralized science is like how you can do science so we have science like it isn't doing science isn't accelerating as quickly as it could or as accurate as it could be and people have thought like if you could decentralize it and remove it from Academia if you could like try and uh not just have it done in universities or in research Labs but sort of anyone could participate in science in any [1:29] anyone could uh be incentivized to do science you could have a lot more Innovation that happens like imagine before GitHub existed or everyone was sort of a professional programmer and then you realize that like oh my God there's so much more talent that exists within the world to oh is any if you guys guys um if there's room everyone take a step to the left and can you can someone in the back open the windows also someone here you you open a window open the [2:11] [Music] window is everyone out of that Corridor is everyone can is everyone is everyone out of that Corridor there is everyone out somebody there's more people outside okay so people keep moving to the left keep going keep going and scooch in the chairs get tighter to each other the love of your life could be in the corridor and you won't even know it keep going keep going we can squish keep going keep [2:54] going okay while while they're doing it um smile smile smile is everyone out now is are people still there sort of keep going keep going if you can squeeze in this way guys if you're sitting take a SK take a take a nudge in come closer to me and you guys too come closer [Music] yeah yeah come closer come closer and guys go go this way go this [3:40] way go this [Music] way come around guys there's a lot of more room all right are there still people organize it are there still people outside because we really don't want people outside no good everyone's inside okay good okay so um okay let's get a group picture just while everyone's quiet panor everyone smile smile smile it's a panoram panorama [4:33] okay okay so um so these are the next talks we have coming up um so we have one on um publishing in Academia so uh Dr muang is an amazing um person she found out at the National Institute of Aging was like um susing some of their fuzzing some of their data and she was the person that actually figured it out initially so if you want that one's going to be um next week 7 days from now and if you want an RSVP here is the QR code to do that cool um then we have who here is like [5:16] into supplements or like longevity or Brian Johnson and you know there's like so many different options of like people that want to like what there's so many different options right and um this guy Dr ctin par is an expert in longevity medicine and you're you're going be in frame you have to it's going to be in the camera um and so no go back go back I like looked at the frame like no all the stand up and like go back okay so um like he is an expert in Long jaity medicine he sold a company on supplements and he's like really really interesting so if you want RSP that one's going to be um just February 16th which is the day after the longevity Summit that Brian Johnson is hosting so [5:58] if you're coming to that the next day we're going to have like a big event with a lot of people there so you guys can be the next ones on that and then how many people have ever dreamt of um like swimming in the Hudson River a lot of people so there's this uh project called the uh the billion oyster project and what they are doing is they are trying to um like clean up large body ways of water and basically make it so that um you can can uh swim there and they figured out that oysters are great filter feeders and can uh basically uh detoxify a lot of the water and also uh [6:40] are carbon sequesters so you can basically uh they they take a lot of carbon out of the water and out of the air and sequestered in their shells and when they die it sort of like keeps it all wrapped up nicely so that's cool and if you want an RSVP that's there on uh March 4th okay um if you guys are excited about science um I started a company which we'll talk about but one of the reasons we Host this event U we oh I want to give a shout out to our sponsors um Finance uh salana uh movement and base they are all very nice chains that sponsor uh D dite NYC so uh give them a round of [Applause] [7:21] applause and uh so if you want to trade coins you can do so on any of those names um but that's just right now so also we started a company called uh or we started a company called Stadium science and we're looking to hire we're hiring for like programmers or uh project managers and stuff like that and how many people are here familiar with like poly Market poly Market yeah and you could like bet on political things that's cool right but what if you could like bet on science like that's that's what that is well so I'll give you a quick brief overview of that so like the old way of doing science was like um you know doing it in universities um and uh you know relying on Experts and in nature or in like in universities and [8:02] the new way of doing and like we saw in the election this time around um like the old way of doing elections is like with with Fox News and CNN and they have the pundits who are very excited about whatever they're talking about but then what people were actually doing is they were going to poly market and saying like oh uh there's actually like we can get insight into who's going to win the election because people are incentivized to um people are incentivized to like talk to their neighbor or come up with new methods of validating the data and so what we're trying to do with uh Stadium science is come up with a new way of validating science and so instead of relying on the experts who are in nature or in universities you can rely on prediction markets and you can solve those prediction markets with like data [8:43] so you can actually like make money by doing science and so the the idea is to make science like profitable and so people who make accurate predictions um will make money off of it so we have a lot of different we we get data from a lot of different sources so we get it from wearables or we get it from Labs um and be trying and like gamify it so some of the competitions right now we're doing competitions on intermittent fasting so people are like uh not sleeping or not eating for a little bit and then we're tracking how it affects their sleep um we're doing a thing on hair with hair D on uh like hair racing where you can accelerate hair with different types of drugs we're doing one on sperm racing where people can subit [9:23] samples and go to a lab like Lab Corp and like see you as the fastest firm but this is something that like Brian Johnson is really into is like sort of gamifying health and like saying that like there's things you can do to affect your health and then we add like a a way for people to um to uh compete on it we're also doing things with like how Vitamin C affects sleep but these are sort of the the intro things the projects that we're working on and creating like markets for each of these um so we have prediction markets on like who's going to sleep the most from like you know you can people that wear an AA ring if you want to sign up you can contribute your data and then you can allow people to on you and you you get a you get a port if you contribute the data you actually earn money through getting a portion of the [10:04] trading fees that we generate um so if you don't want to compete if you don't want to submit your data though you can also bet on other people's data so you don't everyone doesn't have to be a lab mouse some people can be the the other people that are in charge um and um that's about it so this is D NYC we've had a lot of speakers I don't want to be up here forever so I'm going to give it over to our next speaker does any have any questions on D like what D is or on Stadium science how do we give our data to oh go to Stadium go to stadium. science and uh you can sign up for the website oh okay yeah you we're hiring also if you want [10:44] to help build this product cool all right that let let's let's open another window in the back it's kind of hot so this is Dan Dr Dan Dan I don't know oh and also you guys if you we did this thing if anyone is wearing black you should mingle with each other and because you all are desperate and uh so um you guys should talk once we're done here I wanted to make that announcement um so in case you didn't know all right so without further Ado um I'm going to give you Dan Dan is an expert let me let change [11:24] here so what I really liked about Dan is um Dan I mean I've had this like obviously like hinge and like Tinder they just like are sort of random algorithms and like the future is obviously in like having a more in-depth profile on people and like allowing for AI to match People based on their like psychological profile and Dan has built had built something like this so without further Ado Dan uh so before I jump in I wanted to do a quick poll of the audience about a [12:04] question that I'm interested in but the question requires a little bit of setup here so uh this is a uh this is a chart uh from a 2019 paper it's showing the changing ways that people have met uh changing ways that couples have met from 1940 to 2019 when the paper was published and so it's a little blurry here but if you look at the blue line that's if you're look at the blue line that's uh the percentage of couples that have met through friends uh this black line here that's the percentage of couples that have met through family uh this line right here is the percentage of couples that have met in primary or secondary school and of course the big story over [12:45] the past two decades is this red line uh that's going up and to the right over here which is the percentage of couples that have met online or through dating apps and it overtakes every other means by which people have met starting like 2015 yourself and so the question I'm curious about and I don't know if you guys can see this is is this purple line that I'm adding to the very bottom of this chart here which is the percentage of couples that have had arranged marriages via AI so obviously right now right now obviously as far as I'm aware this is 0% I'm not aware of any couples that have had arranged marriages AI but uh my question here is it's 2025 [13:26] right now within the next 20 years by 2045 will the major will that purple line that I just showed will it overtake all the other lines will the majority of new marriages be arranged by I so just by a show of hands how many people think a majority of marriages will be arranged by AI don't be shy okay I'm going to call that uh maybe 10% question uh let's leave questions for the end there's some ambiguity there but take as div in here um okay so just for some context for why I'm giving this talk so in 2021 to 22 2022 I was running a dating service and I ended up [14:07] organizing nearly 3,000 in-person dates um I was running speed dating events not for the sake of running speed dating events per se it wasn't U um an end in and of itself but it was a means to rapidly collect the real world training data and I was training uh romantic compatibility models off of this and then I built an automated matchmaking system and so this talk is structured in four main parts in the first part I'll talk about how you can use speed dating events to rapidly collect Real World Training data then I'll talk about some interesting insights into the data I'll talk about uh attractiveness how much do [14:48] looks actually matter we'll go into some political and uh religious and ethn ethnicity data um then I'll talk about uh if it's possible to train machine learning models to predict the compatibility between a man and a woman before they've even met and finally I'll talk about uh this automated match making system I built and at the very end I'm happy to answer any questions and also uh I've open sourced all of the anonymized data that I collected so uh I'll share links to that and all the writing uh from the stuff that I published online as well okay so first up um how uh is there someone who's buy the lights that can turn the lights off [15:29] no does anyone buy a light switch Oh that's oh turn the do the opposite the opposite keep some of the lights on in the back yeah that's and there's can someone let someone in at the door there's the pizza there is it not it was like the first there we go Okay cool so first section here is how you might use uh speeding events to rapidly collect Real World Training data so this is a picture from a speed dating event that I ran at a brewery in Austin Texas and for people that are unfamiliar with what the format is generally speaking for heterosexual speed dating [16:11] events women are typically seated at one spot throughout the duration of the event the event might last like an hour and a half to two hours and then men are initially assigned a starting position but then uh while running these events I would walk around with a bell and every eight or nine minutes I would ring the bell and the men would know to rotate up one seat and so by the end of the event everyone gets a chance to meet each other um all of the logistics and the uh data collection was done through uh mobile apps that I built so people would download an iOS or Android app they would create their account they would fill out their profile they'd be able to browse events speed dating events in their area that I was running you know find out details about those events [16:51] purchase tickets and then after the event was over they would be uh swiping left and right on the people they talk to to kind of know whether or not they liked the person that they spoke to and if both people uh liked each other uh I would share contact details in the app and and they could take it from there um so I was experimenting with running speed dating events of different sizes and different age groups and I found that in terms of attendance uh that 20 person Mark was kind of the goldilock Zone in terms of attendance if uh I found that if if you run events that are much larger than this in size if you go up to 30 or 40 people the the big downside here is that it starts to get a [17:32] little bit overwhelming you can imagine if you're going on these speed dates with like 15 or 20 people back to back to back it kind of gets draining unless you're super extroverted so that's the big downside there if the event is much smaller in size if it's say like 10 people uh then the big downside is uh you're not super likely to meet someone that you're compatible with so 20 person 20 person events are kind of the the goalie lock Zone here and uh I was running heterosexual events uh average size was 20 and I made sure to take the uh gender ratio very seriously so the average gender ratio is split evenly between men and women 10 and 10 and the way that I maintain this even gender ratio for every event is by manipulating [18:13] ticket prices and AD spend um so the average ticket price for one of these speed ating events uh at least in Texas was $25 but there was a big range there so in some cases I'm giving away tickets for free in other cases I'm selling tickets as high as $41 and just depends on the gender ratio so as a general rule of thumb uh for the speed daying events that I ran for older age groups it was generally speaking easier to sell tickets to women much more difficult to sell tickets to men and for younger age group let's say for people in their 20s it was easier to sell tickets to men more difficult to sell tickets uh to women there were two types of data that [18:53] I was collecting from each event so the first type of data is the outcome of every date and the second type of data are the attributes and preferences of each participant so let's talk about the outcomes first so speed dating has this Advantage which is that with a comparatively small amount of people you can actually get a large amount of uh real world data so again average size of these events is 20 people evenly split between men and women and even with just 20 people you actually at the end of the event get to observe uh the outcomes of a hundred possible relationships right because everyone needs each other and this is what the data of the outcomes look like from each of these events [19:34] pretty straightforward every row here represents a date uh so for example if we look at the first row uh guy a is going on a date with a woman D he did not like her she liked him so there was no mutual interest as opposed to this last date here seeing G and going on a date they both like each other which means there is mutual interest which means contact information is going to be shared in the app um the second type of data I was collecting was the attributes and preferences of everyone who was in attendance and there were kind of two flavors of this um the first type of data that I was collecting about each guest was this 100 question compatibility survey and I'll go into more detail into this in a second but it [20:15] basically basically covered uh your personality traits your intelligence your identity questions around health and lifestyle and a bunch of miscellaneous questions as well and the important thing to note here is that this was actually a requirement to purchase ticket so when people are in the app browsing for different speed dating events I didn't allow anybody to just purchase a ticket the purchase was gated upon successful completion of this compatibility survey and then the second type of data that I was trying to collect on each person in attendance was uh a rating of their physical attractiveness and basically the gist here is people would upload pictures and then I would get them rated by opposite sex um uh like independent I would get independent ratings by the opposite sex [20:57] uh to get ratings for physical attract and I'll talk about this in more detail in a second too but first I'll talk about this compatibility survey um so one thing that I was trying to get a sense of is your personality and I was using this five Factor model of personality from like a Psy 101 class and so there are five factors here uh one of them is openness to experience this is a measure of like your curiosity creativity preference for novelty and variety conscientiousness is this tendency to be organized Dependable self disciplined extroversion most people are familiar with agreeableness is this concern for social harmony tendency to get along with others compromise and [21:37] neuroticism it says emotional stability here it's actually the inverse of emotional stability it's this tendency to feel uh negative emotions like anxiety or depression um so this is a screenshot from the app this is actually a little bit outdated it was before I Whittle down the survey to 100 questions it used to be a little bit more but uh this question was uh trying to get at this open experience uh uh factor of personality so a good percentage of the 100 questions on this survey were dedicated to personality traits the second type of data that I was trying to get in this um uh yeah so I was trying to get a [22:17] sense of people's intelligence and it turns out that depending on the study that you look at sap scores correlate really well with General measures of IQ somewhere in the range of 7 to9 so so basically what I did is find uh super old like 20-year-old SAT exams took a bunch of random questions this is not literally one of those on there but uh some of the questions were literally sat questions um bunch of other things in the survey uh questions around identities so how users identify in key areas of the lives including political beliefs uh religious affiliations and ethnic backgrounds it also assessed how much these factors mattered in the romantic and social relationship so for [22:58] example I ask questions like how important is it that your romantic partner shes your religious beliefs how comfortable are you being friends with someone that disagrees with you on important political topics there are a lot of questions around health and lifestyle like alcohol consumption drug use exercise exercise habits and a bunch of miscellaneous questions around your income height age location and how many children you want and so on and so forth so that's the 100 question compatibility survey the second type of data I was trying to get a sense of is how attractive you are physically and so people would uh upload a picture of themselves in the app and then I would recruit opposite sex workers from the service called prolific to rate their pictures and the goal being to get at [23:39] least 10 independent ratings of attractiveness for each person in order to get an average and for those people that don't know prolific is kind of similar to Mechanical Turk um people use it in academic settings as far as I'm aware uh and you can kind of get people to complete tasks for you from like embedded and customizable demographic group so had a good experience working with them and I'd recommend working with them if uh you need something like this um so the people that were rating the pictures were direct to this little hacky tool that I built uh where you would rate uh people's pictures so in this case we're rating Elmo's pictures the instructions here are pretty straightforward please rate the physical attractiveness of the person in a picture on a 1 through 10 scale with one [24:22] being least attractive five being average and 10 being most attractive so once you've got all this data you got the physical attractions data and you've got the responses to the compatibility surveys these were distilled into a comprehensive set of user attributes and preferences uh and an example of this is is on the right here this is like a set of user features so for example you know you've got information like the user's age uh who they're attracted to what percentage of those sat questions they get correct scores on personality their agreeableness conscientiousness extrovert and of openness what is their average physical attractiveness rating how tall are they how much money do they [25:02] make what's their ethnicity religion politics how much like their drug use and so on and so forth so that's a little bit about how you might use uh speed dating to rapidly collect Real World Training data uh now I'll talk a little bit about some cool insights into this data and I'll give like a hopefully this is not necessary I'll I'll give some disclaimers a little Preamble here uh some people get a little bit sensitive about this data I'm about to make some general ations about groups I'm about to talk group about group averages uh that misses a lot of nuance obviously uh it's a generalization there's exceptions to the rule and uh obviously knowing something about a group average has no bearing or implication on any individual [25:44] person okay so let's talk about physical TR okay so this is a uh distribution of uh women's uh physical attractiveness in the eyes of men and what you'll notice here is that it's a pretty normal looking distribution it looks kind of like a bell curb where uh in the eyes of men most women are kind of average looking the bulk of women are in kind of in the middle very few women that are incredibly attractive very few women that are incredibly unattractive and so actually the median rating of men of women by men is right down the middle is a 5.0 out of 10 so takeaway here men [26:27] perceive the median woman to be average looking let's look at the same same data same chart I should say uh but this time of men's physical attractivness ratings in the eyes of women and what you'll see here is it's a left view distribution the majority of men are below average looking in the eyes of women and in fact women rate 72% of men as the low average in physical attrac so this this difference in this is the same exact data that I just showed it's just easier to see the difference uh on the same axes um and so I've written a [27:07] little bit about this online and there's often certain comments uh I hear like um you know uh women are unfairly judging men's attractiveness or women should be uh grading men on a normal distribution just like men do for women and I think that's not necessarily the right takeaway here um like there's no law of nature that says everything must be normally distributed there's there's many examples in nature of things that are not normally distributed and I think there's good evolutionary reasons for why women might have a higher threshold for attractiveness and it basically comes down to this idea that um sperm is cheap but eggs are expensive [27:49] um so yeah I think there there might be some good reasons for why women might be a little bit pickier in certain contexts on average okay so uh men are less attractive than women on average but how much does this actually matter in dating how much do looks matter so um speak up speak up uh looks matter for everyone um and this uh this data here is is I basically took men and women and I grouped them into four buckets uh from least attractive to most attractive quartiles right least attractive second quartile third quartile most attractive for both men and and women that's on the x- axis [28:30] and on the Y AIS you're seeing the fraction of the opposite sex that like them at a speed dating event so after having an 8 minute long in-person date and so what you see here is this up into the right relationship which which suggests that the more attractive you are the more the opposite sex is going to like you after a short date um so perhaps un unsurprising for both men and women you see this up into the right relationship looks matter for and for women but if you look at the slopes here um looks matter more for women than men and you can see this by comparing the uh difference between the least attractive and the most attractive quartiles so [29:12] there's a 50 percentage Point jump in likability between the least attractive and most attractive quartiles for women whereas for men this increase in likability exists it's positive but it's less than half it's only 21% of points so looks matter for everyone but at least in a dating context they matter more for women uh than for men uh what's the relationship between age and attractiveness uh perhaps unsurprisingly older people are less attractive on average um so on the x- axis here you uh it's it's men and women broken up by decade 20s 30s and 40s and on the Y AIS here you've got their [29:53] physical attractiveness ratings and as opposed to the up and to the right relationship we were looking at for the previous charts in this case we see a down into the right relationship which suggests that as people get older they're less attractive and this is the case for both women and men but again if we look at the uh slopes here there's a difference uh the decline between in physical attractiveness for women in their 20s to their 40s is 1.2 points in in in attractiveness whereas for men it's only half a point so older people are less attractive but the decline is steeper for women okay let's talk about politics religion and ethnicity so one question I [30:34] was asking on the compatibility survey is um how important is your partner's politics religion or ethnicity and the possible answers you could give range from not important at all to very important and so this chart is showing you the percentage of people that said it was somewhat or very important uh that their how important uh their Partners politics religion ethnicity were and the percentage of people that saying it was somewhat or very important so if you look at the politics uh line here something like 70% of people are saying that's important to them versus ethnicity is like 30% so I thought this was interesting because I feel like if you were to ask this question a 100 years ago you would get a very different [31:16] ordering of these results um if we look at the same exact uh kind of chart here but this time broken up by gender what we see here is that in each case for politics religion and ethnicity women care more than men about their about their partner um and so I think this goes back again to in certain contexts women uh are a little bit bakier okay A few rapid fire charts on sex drugs and spirituality uh women are less interested in casual sex perhaps unsurprising uh men are more likely to use drugs this one was interesting women are eight times more likely to be bisex [32:00] um and this is interesting uh you know if you ask if you ask four different people who say they're spiritual what they mean by that it's a kind of a wishy-washy vague term I never know what anyone means when they say they're spiritual um in some you know in some context I might even say that I'm spiritual uh but what's clear in the data is that spiritual people seem to use marijuana more and have more sexual partners and then last thing here there's a little tool that I built which is available at dat filters.com and and I'll share a link to this uh at the end of the talk as well but this is a tool that shows you what percent of singles at least according to my data set have [32:40] the traits you might care about so you can filter uh you can switch this between men and women and you can filter the population of single men or single women by any criteria you care about so if you want to find that 65 guy with blue eyes and fin you might try try okay so that's some insights into the data that I collected now let's talk about whether or not it's possible to train machine learning models to predict the romantic compatibility between uh a man and women before they've even met or to say this a little bit more concretely so given all these user features given everything you know about the man and wom given what you know about their personality their IQ their [33:21] physical attracts rating and so on and so forth can you use this data can you use these features to predict of these outcomes of these dates and so there's different outcomes you could think about uh predicting I was most interested in the mutual interest uh outcome which means whether both the man and the woman like each other after that first date so there's this concept in evolutionary psychology called assortative mating and assortative mating basically is this idea that birds of a feather flop together and I stole Wikipedia's description of this here so it says human mating is inherently nonrandom despite the common Trope Opposites Attract humans generally [34:03] prefer mates who share the same or similar traits and I think most people have like an intuition around this so uh it tends to be the case that attractive people tend to date other attractive people smart people tend to date other smart people but this extends uh to many more Dimensions beyond that and so in order to try to capture this in the model in addition to the individual traits for the man and the woman I also introduced these assortative M mating features which captured the similarity or dissimilarity between the man and the woman so for example you have the man's average physical attract rating you have the woman's average physical attract rating and you can compute the difference here and you have this [34:43] average physical attractiv rating diff where diff stands for difference and you can do this for IQ and height and so on and so forth this is what the uh training and test data set look like um so every row here represents a date and then you've got the individual features for both a man and a woman uh as well as the assortative mating features the differences between the woman and man's age IQ personality and so on and so forth and then finally on the very right you've got that Target variable that variable that you're trying to predict in this case it's mutal interest whether both a man and a woman uh ended up liking each other and so the way that this typically works is [35:25] um when you're you're training a model you randomly split this data set into two groups a training set and a testing set and while training the model uh it has visibility into the outcomes it can see what happened on that date and it's learning from that from that data when you're evaluating the model you want to see how accurate it is you hide that from the model and instead you ask the model to make predictions as to what it thinks is going to happen in this case it spits out um probabilities for mutual interest this p mutual interest column so as an example this first row a date between A and B the model says very high likelihood is 73% Chance in This toy [36:06] example that uh they're going to uh find each other interesting and want to meet for a second date whereas in the second example between C and D very low probability they're probably not going to get along uh so the analogy here is kind of to like a student studying for an exam when they're doing homework problems they can look at the back of the book they can see you know did I do thisk problem right they can look at the answer key when you want to evaluate the students's knowledge when you want to see if they absorb the material uh you have them take exam and obviously during the exam they can't look at the answer key but you can evaluate that exam later on by comparing uh what answers they got right or wrong okay so how well does this [36:47] compatibility model do so this chart is showing the probability of at least one date with mutual interest and it's comparing the model's performance which this purple solid line up here to the dotted Gray Line which is uh representing the base rate and I'll explain what this base rate is first so if you imagine taking a random man and a random woman and pairing them up on a date the likelihood that there's going to be mutual interest at the end of that date is 12% um and the what the dotted gry line is showing you is that assuming every data is independent if you keep making these random pairings and this person is going on uh more than one date um the [37:29] probability of at least one of those dates being successful of one of those dates having mutual interest is going to be increasing right uh assuming each one is independent and so that's why you see that up into the right relationship but what if instead of doing this random pairing you use the model to select uh a date for this person that it ranks in the top fifth percentile of scores so you're now being selective about what you choose and you're using the models scores to make the selection well then the probility of mutual interest uh from a first date jumps up to 45% and this is another point of uh so this is a Forex increase in the roughly Forex increase in the probability of a successful date [38:09] as compared to random chance um just for another point of comparison here uh I won't mention them by name but uh one of the most well-known Nationwide matchmakers uh in the United States which you know these matchmakers uh charge thousands and thousands of dollars for their Services they're very Cy about their metrics basically because they do a not a great job um and if you look at their data you can try to figure out like how what's their performance that that would be equivalent to this kind of metric I'm showing you guys here and in the most optimistic case it's 35% so my little hacky toy thing from the speed inting thing is outperforming the most well-known Nationwide Matchmaker uh and [38:51] I think this is basically just because it's difficult to scale human matchmaking scaling human intuition is kind of difficult okay so what were the most important features out of the model there are 118 total what are the features in this model that are most predictive of whether or not there's going to be mutual interest between a man and a woman so this is a chart showing the top 10 features in this random forest model uh in descending order of importance and they're colorcoded by category here so the blue ones are personality related features the kind of pink one are attractiveness related features and I'll just make two quick observations here so out of these top 10 features eight of them are [39:32] assorted of mating features uh and you can tell because there's a suffix undor diff those are all uh those assorted mating features which are trying to capture the difference or the similarity between the man and the woman on the date and then the most the single most important feature in the model is the difference between the man and the woman's physical attractiveness ratings but what I thought was interesting is that coming in a very close second is this difference in neuroticism um and neuroticism again is this inverse of emotional stability it's this tendency to feel anxiety and depression so I thought that was pretty interesting are those positive or negative loads let's do uh questions at the end if that's okay um [40:14] uh so this is this is a hacky kind of prototype model um in the grand scheme of things like uh it's it's trained on a very small amount of data um there's many uh ways you can think about improving it um some ideas of like interesting data you could collect that could would be interesting to see how it would perform um you you could try to capture these intangibles is it possible to capture these wispy qualities like communication ability sense of humor Vibes the attractiveness of someone's voice um my 100 question compatibility survey was all self-reported data right and self-reported data can be unreliable so how would you handle exaggeration or [40:55] dishonesty I also didn't includ any education I didn't collect any Education data and then uh in terms of the outcomes uh if you imagine uh relationships like almost like a marketing funnel um you know the outcomes I'm capturing are very shallow in that funnel it's just it's just a first date uh a very short first date kind of outcome and it would be interesting to capture outcomes deeper into that funnel so is the couple still dating after three months uh did they get married how happy are they with their relationship so that's a little bit about uh building a compatibility model um so uh I use [41:36] this compatibility model to build this automated matchmaking service and I'll talk a little bit about how that works so the first two steps here very similar to what I've already talked about before first step is um person creates their account they complete their compatibility survey upload their their photos from this data you extract features you get their photos rated you distill all these these features into into a comprehensive set of attributes and preferences The Next Step here is uh for any given person that's interested interested in um uh getting set up you compute the set of all possible dates they could be going on and what I mean [42:16] by possible dates is it's dates that they haven't been on but they could be going on and for each one of these pairings you use your romantic compatibility model to uh estimate the likely of mutual interest for each one of these possible dates so it's spitting out uh a probability of mutual interest for each one of these pairings the next upep here was around eligibility and basically the thinking here was um I didn't want uh I I I did not I restricted um uh I did not allow everyone to purchase a matchmaking package uh I only wanted to allow a subset of people to purchase a ma matchmaking package and only if I felt [42:58] that um I had reasonable confidence and reasonable evidence that I could that it would be worth their time and money to go on a on a date with someone that I set them up with and so this is kind of how that eligibility process process worked so first you start with this universe of all possible dates that person could go on and you progressively start applying different filters and if at the end of that filtering process you're left with a reasonable number of quality dates then they would be allowed to purchase a date otherwise they would be put on a weight list so the first set of filters here were these heuristic based filters um people generally speaking have these hard constraints in terms of who they're willing to date [43:38] they usually are not interested in like a longdistance relationship so there's hard filters on location most people have age criteria right uh a chart that I showed previously was saying that you know 70% of people care about politics at least in 2021 2022 and so some people are just incompatible politically um filter people out by height as well so first you start with that set of possible dates you use these hard constraints and filter people out then you apply these statistics based constraints where you filter out people if their probability of mutual interest is below a threshold and if at the end that all at the end of that whole process um there's still uh a minimum [44:20] number of quality dates then you allow that person uh to purchase a Matchmaker package and then finally if they purchase a package you uh help uh arrange a date for them and so the way this works is you start going down that list of uh in descending order by probability of mutual interest reaching out to that person asking hey are you interested in being set up on this date with this person if they say yes that's great you set them up on the date if not you move on to the next person on the list okay so how did it go uh so I was running a trial in Austin Texas uh where I was charging $129 to be set up on a single date and at the [45:00] end of this whole eligibility process for this trial there were 25 people that were eligible that I felt like okay I feel like I have some confidence in being able to set you up on someone uh where there's a reasonable chance you guys would get along out of those 25 people one person uh in this case a guy ended up purchasing it so a 4% conversion in there um so I was started trying to set this guy up on a date and I started I started reaching out to the women uh that uh would could be compatible with him and I'm either I'm getting two flavors of responses one is no response at all um the second response is oh man I feel bad I signed up for your app you know nine months ago [45:40] or six months ago I'm now in a new relationship no longer interested so very disappointing was not able to set this guy up on a match and I had to refund him for his purchase so a little bit disappointing um there's lots of things you could think about that I could improve right the model could be more accurate the design of the app could be better the level of service could be better there's millions of things that could have been better but ultimately I did not fail because of the tech or the product or the design I fail because of distribution um even if you have uh a perfectly accurate and omniscient matchmaking model it's not super useful if a customer soulmate is not an active and engaged uh user on your [46:23] service and so just to illustrate this point um of distribution um if you're thinking about maybe building a matchmaking service or a dating service like this it's useful to go through this uh thought experiment which is how many users do you need to have to have high confidence of setting up one successful date so let's say a woman comes to you she's interested in being set up on a date and there's all these guys that are within like a 30 mile radius which is the maximum she's willing to date um within uh and again this is just a thought experiment these are just back of the envelope numbers you can plug in your own here this is just for illustration but you know immediately there's all these hard constraints and we're going to filter out 30% of the [47:04] people so 70% of these guys are left right you filter them out because of uh political mismatches or age or height um then you score the remaining guys with your compatibility model right and you're going to filter out 95% of these people uh because we only want really high quality matches but you can plug your plug in your own numbers here right then you're left with these guys here who in theory uh could be good people to go on a date with and so you start reaching out to them hey do you want to be set up with this woman but for all sorts of reasons 80% of them are not going to show up they uh you know may have they may have gotten super busy with work they may have mov cities they may have signed up nine months ago now [47:45] they're in a your relationship right and then we know uh from the model there's like basically a coin flip as to whether there's going to be mutual interest or finally okay this is the guy that there was some connection with after that first date so if you plug in all these numbers feel free to plug in your own the the takeaway here is that you need a at minimum hundreds of customers uh hundreds of uh of the of members of the opposite sex within any um you know certain radius of a a person that might be interested in your service so this is was not entirely surprised the fact that distribution is the overwhelming reason for the graveyard of dating startups it's not a [48:26] close guarded secret I had known all this going in and yet it was too easy to get distracted instead of obsessively focusing on the most critical element for a successful dating service my attention drifted to all sorts of different places um and while I abstractly understood that those were mistakes at the time I grasp their significance in a much more visceral and palit way after going through this so if you're interested in in doing something like this if you're interested in starting a dating app or a matchmaking service it's worth I guess this this would be my t are for you which is it's comparatively very easy to figure out why existing dating services are not great and how could they and how they could be improved everyone and their mom [49:07] and their dog has an idea for a dating uh app but it is comparatively much more difficult to figure out how to efficiently acquire a dense local network of active and engaged users and it seems to me that until you have a clear solution to that exist existential challenge any investment into a new dating service seems a bit misguided if your goal is to build a successful business um I still think this is a problem we're solving because choosing a spouse is arguably the most important decision one ever makes um if you think about how much money and time people spend on their choice of where they go to college um there's so much anxiety for many people uh around this in high [49:49] school um people you know the sticker price for many universities is 60k right now right people will invest a quarter million of doll quarter million dollars into where they're going to school people will uproot their lives and their family family's lives to move cities for the sake of a career or a job but arguably the person that you choose to marry is more important than the university you went to or any given job so I think this is a problem we're solving I was not able to be successful here I think someone will be and I think that within the next 20 years uh someone will figure this out and the majority of marriages are going to be arranged by AI uh this is a QR code that will take you to all the links and the data that I [50:30] talk about I'm happy to answer any questions as [Music] well if someone could turn the lights on turn the lights on yeah um so you have that list of features by ranked by feature importance but did you also look at whether each of those respective features were positive or negative loading I did not no that would be interesting because it could actually be like you want someone way more neurotic than you right if you're just looking at yeah the relationship there would be very interesting okay yeah hey hey Dan Sasha here first thank you for sharing that Journey with us the [51:11] second question is well common is what about the conversation can you talk to us about like did was there a pretty standard format was there not cuz 9 minutes isn't a lot of time to get a sense for a Vibe like I'd imagine if you recorded and I hope you didn't but like if someone had laughed a bunch of times maybe they would be compatible we've all been in a situation where a conversation goes better well with someone that then you think like how did you think about that in terms of both setting the context there and generating I I didn't I mean that that could be interesting to think about designing an event where you structure the conversation more yeah in the back there I feels if you're in the back and you're speaking quiet oh I'm sorry no go go I'm a Matchmaker and I [51:55] think the most important I think a lot of yeah totally I I did not work at all on any coaching products yeah hi um I have a more deck question like did you experiment with different mod or like how did you chose different different what the mod did you like experiment with or how did you uh no I didn't experiment with any other ones yeah all all this data is public too so if you want to train your own model uh you can do that yeah yeah so it [52:36] seems like um there have been other attempts to have you know better um that have tried to incentivize the successful creation of the match and they've mostly gone